Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: bring the picker's Connected rows up to the On Device rows (#11196)

Michael Hanb902425d7944
diff --git a/studio/backend/core/inference/provider_model_capabilities.py b/studio/backend/core/inference/provider_model_capabilities.py
index 9932251f3ca..a4f8d75243b 100644
--- a/studio/backend/core/inference/provider_model_capabilities.py
+++ b/studio/backend/core/inference/provider_model_capabilities.py
@@ -114,6 +114,13 @@ def _trim_models_dev_model(model: dict[str, Any]) -> dict[str, Any]:
inputs = modalities.get("input") if isinstance(modalities, dict) else None
if isinstance(inputs, list) and inputs:
entry["input"] = [m for m in inputs if isinstance(m, str)]
+ # Total context window. The frontend merges this payload over its bundled snapshot per model,
+ # so a field missing here erases the bundled one. limit.input is the prompt share of the
+ # window and limit.output the completion cap, which max_output_tokens already carries.
+ limit = model.get("limit")
+ context = limit.get("context") if isinstance(limit, dict) else None
+ if isinstance(context, int) and not isinstance(context, bool) and context > 0:
+ entry["context"] = context
return entry
diff --git a/studio/backend/tests/test_provider_model_capabilities.py b/studio/backend/tests/test_provider_model_capabilities.py
index a9de3483f9a..1cf72532909 100644
--- a/studio/backend/tests/test_provider_model_capabilities.py
+++ b/studio/backend/tests/test_provider_model_capabilities.py
@@ -298,6 +298,59 @@ def test_models_dev_catalog_is_trimmed_and_remapped():
assert catalog["gemini"]["gemini-9-flash"] == {"reasoning": True, "input": ["text", "image"]}
+#: `limit` shapes that must and must not yield a context window. The frontend merges this payload
+#: over its bundled snapshot per model, so emitting a junk value is worse than emitting none: it
+#: displaces a good bundled window rather than falling back to it.
+_CONTEXT_CASES = [
+ ({"context": 1_050_000, "output": 100_000}, 1_050_000),
+ ({"context": 480}, 480),
+ (None, None),
+ ("not-a-dict", None),
+ ({"output": 4096}, None),
+ ({"context": None}, None),
+ ({"context": 0}, None),
+ ({"context": -1}, None),
+ # bool is an int in Python, and True > 0, so without the explicit bool check this would
+ # emit `"context": True` for any provider that sends a flag here.
+ ({"context": True}, None),
+ ({"context": False}, None),
+ ({"context": 1000.5}, None),
+ ({"context": "1000"}, None),
+]
+
+
[email protected](("limit", "expected"), _CONTEXT_CASES)
+def test_the_published_context_window_survives_trimming(limit, expected):
+ from core.inference.provider_model_capabilities import _trim_models_dev_model
+
+ model = {"modalities": {"input": ["text"]}}
+ if limit is not None:
+ model["limit"] = limit
+ entry = _trim_models_dev_model(model)
+ assert entry.get("context") == expected
+
+
+def test_the_context_window_reaches_the_served_catalog():
+ from core.inference.provider_model_capabilities import trim_models_dev_catalog
+
+ raw = {
+ "openrouter": {
+ "models": {
+ "openai/gpt-4o": {
+ "modalities": {"input": ["text", "image"]},
+ "limit": {"context": 128_000, "input": 120_000, "output": 16_384},
+ }
+ }
+ }
+ }
+ catalog = trim_models_dev_catalog(raw)
+ # limit.input is the prompt share and limit.output the completion cap; neither is the window.
+ assert catalog["openrouter"]["openai/gpt-4o"] == {
+ "input": ["text", "image"],
+ "context": 128_000,
+ }
+
+
class _FakeCatalogClient:
calls = 0
fail = False
diff --git a/studio/frontend/scripts/refresh-model-catalog-snapshot.mjs b/studio/frontend/scripts/refresh-model-catalog-snapshot.mjs
index 1a97c0b8084..e856ede39b6 100644
--- a/studio/frontend/scripts/refresh-model-catalog-snapshot.mjs
+++ b/studio/frontend/scripts/refresh-model-catalog-snapshot.mjs
@@ -47,6 +47,9 @@ function trimModel(model) {
if (options.some((option) => option?.type === "toggle")) entry.toggle = true;
const input = model.modalities?.input;
if (Array.isArray(input) && input.length > 0) entry.input = input;
+ // Total context window. limit.input is the prompt share, limit.output the completion cap.
+ const context = model.limit?.context;
+ if (typeof context === "number" && context > 0) entry.context = context;
return entry;
}
@@ -80,6 +83,8 @@ const lines = [
" efforts?: readonly string[];",
" toggle?: true;",
" input?: readonly string[];",
+ " /** Published context window in tokens. */",
+ " context?: number;",
"}",
"",
"export const MODEL_CATALOG_SNAPSHOT: Readonly<",
diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index af6b9d99e88..2ae839a7058 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -21,6 +21,8 @@ import {
SidebarModelConfig,
useActiveModelConfig,
useModelConfigHandoffStore,
+ pinnedReasoningEffort,
+ useModelReasoningEffortStore,
} from "@/features/model-picker";
import { ProjectComposer, Thread } from "@/components/assistant-ui/thread";
import { usePlatformStore } from "@/config/env";
@@ -182,7 +184,7 @@ import {
getTrainingCompareHandoff,
} from "./lib/training-compare-handoff";
import {
- clampReasoningEffortToLevels,
+ externalReasoningTakesEffort,
getExternalReasoningCapabilities,
getProviderCapabilities,
modelCatalogVersion,
@@ -193,6 +195,7 @@ import {
providerSupportsBuiltinWebSearch,
providerSupportsFastMode,
reasoningFieldsAfterCatalogRefresh,
+ resolveExternalReasoningEffort,
subscribeModelCatalog,
} from "./provider-capabilities";
import {
@@ -220,7 +223,11 @@ import {
CHAT_WEB_FETCH_TOOLS_ENABLED_KEY,
PENDING_CHAT_ATTACHMENT_KEY,
loadOptionalBool,
+ noteEffortDisplacedByPin,
+ pinHoldsLiveEffort,
readPendingAttachmentTargetClaim,
+ reconcilePinnedReasoningEffort,
+ takeEffortDisplacedByPin,
threadScopedOverride,
useChatRuntimeStore,
} from "./stores/chat-runtime-store";
@@ -2538,41 +2545,28 @@ export function ChatPage({
},
);
const state = useChatRuntimeStore.getState();
- const preferredEffort = state.reasoningEffort;
const effortLevels = reasoningCaps.reasoningEffortLevels;
- const clampedEffort = clampReasoningEffortToLevels(
- preferredEffort,
- effortLevels,
- );
- // Per-provider default effort: Anthropic gets the highest level, since Claude's adaptive thinking
- // adjusts cost per turn; OpenAI gets "high"; everyone else "medium". Overridable via Think.
- const isAnthropic = provider?.providerType === "anthropic";
- const isOpenAI = provider?.providerType === "openai";
- const anthropicTopEffort = effortLevels.includes("xhigh")
- ? "xhigh"
- : effortLevels.includes("high")
- ? "high"
- : clampedEffort;
- const openaiDefaultEffort = effortLevels.includes("high")
- ? "high"
- : effortLevels.includes("medium")
- ? "medium"
- : clampedEffort;
- const catalogDefaultEffort =
- reasoningCaps.defaultEffort &&
- effortLevels.includes(reasoningCaps.defaultEffort)
- ? reasoningCaps.defaultEffort
- : null;
- const nextReasoningEffort = reasoningCaps.supportsReasoning
- ? (catalogDefaultEffort ??
- (isAnthropic
- ? anthropicTopEffort
- : isOpenAI
- ? openaiDefaultEffort
- : effortLevels.includes("medium")
- ? "medium"
- : clampedEffort))
- : state.reasoningEffort;
+ // Through the shared resolver, pin included: this runs on reload and on every provider
+ // resync, and resolving it without the pin is what put the provider default back over a
+ // level the user had set on the model's row.
+ const pinnedEffort = externalReasoningTakesEffort(reasoningCaps)
+ ? pinnedReasoningEffort(inferenceParams.checkpoint, effortLevels)
+ : null;
+ const nextReasoningEffort = resolveExternalReasoningEffort({
+ caps: reasoningCaps,
+ providerType: provider?.providerType,
+ // Same as the switch below: a checkpoint that changed without going through it leaves the
+ // previous model's pin in the live level, which an unpinned model must not inherit.
+ current:
+ !pinnedEffort && pinHoldsLiveEffort()
+ ? (takeEffortDisplacedByPin() ?? state.reasoningEffort)
+ : state.reasoningEffort,
+ pinned: pinnedEffort,
+ });
+ // The chat's own level, before the pin takes its place, so clearing the pin can put it back.
+ if (pinnedEffort && nextReasoningEffort !== state.reasoningEffort) {
+ noteEffortDisplacedByPin(state.reasoningEffort);
+ }
const supportsBuiltinWebSearch = providerSupportsBuiltinWebSearch(
provider?.providerType,
selection.modelId,
@@ -2668,6 +2662,36 @@ export function ChatPage({
// Reruns once settings hydrate: this normalization reads the stored pills and clamps them to the
// model, and hydration refreshes what it reads, so it has to be applied last.
}, [externalProvidersForChat, inferenceParams.checkpoint, settingsHydrated]);
+ // Another tab can change the active model's pin (the pin store listens for the storage event),
+ // and the normalization above reads the pin through a getState helper without subscribing, so
+ // the composer kept the old level until a switch. Only the effort here: the pills and the rest
+ // of that block are not this one's to rerun.
+ const activePinnedEffort = useModelReasoningEffortStore(
+ (state) => state.effortByModel[inferenceParams.checkpoint],
+ );
+ const appliedPinnedEffort = useRef(activePinnedEffort);
+ useEffect(() => {
+ if (appliedPinnedEffort.current === activePinnedEffort) return;
+ appliedPinnedEffort.current = activePinnedEffort;
+ const selection = parseExternalModelId(inferenceParams.checkpoint);
+ if (!selection) return;
+ const provider = externalProvidersForChat.find(
+ (p) => p.id === selection.providerId,
+ );
+ const caps = getExternalReasoningCapabilities(
+ provider?.providerType,
+ selection.modelId,
+ {
+ isReasoningProvider: provider?.isReasoningModel === true,
+ baseUrl: provider?.baseUrl ?? null,
+ },
+ );
+ reconcilePinnedReasoningEffort({
+ checkpoint: inferenceParams.checkpoint,
+ caps,
+ providerType: provider?.providerType,
+ });
+ }, [activePinnedEffort, externalProvidersForChat, inferenceParams.checkpoint]);
// A catalog that lands after selection refreshes only the stored reasoning fields (the effort shortcut reads them),
// never the selection defaults above, so a chosen effort and the pills survive the refresh.
const modelCatalogChange = useSyncExternalStore(
@@ -2695,6 +2719,13 @@ export function ChatPage({
useChatRuntimeStore.setState(
reasoningFieldsAfterCatalogRefresh(useChatRuntimeStore.getState(), caps),
);
+ // After the levels, not before: a pin is only applied while the catalogue calls it legal, so
+ // the refresh that publishes the level is what puts the model's own pin in force.
+ reconcilePinnedReasoningEffort({
+ checkpoint: inferenceParams.checkpoint,
+ caps,
+ providerType: provider?.providerType,
+ });
}, [modelCatalogChange, inferenceParams.checkpoint]);
const canCompare = useMemo(() => {
return Boolean(inferenceParams.checkpoint) && !isExternalModel;
@@ -3190,41 +3221,25 @@ export function ChatPage({
baseUrl: selectedProvider?.baseUrl ?? null,
},
);
- const preferredEffort = store.reasoningEffort;
const effortLevels = reasoningCaps.reasoningEffortLevels;
- const clampedEffort = clampReasoningEffortToLevels(
- preferredEffort,
- effortLevels,
- );
- // Same per-provider default policy as the useEffect above: Anthropic highest level, OpenAI
- // "high", everyone else "medium".
- const isAnthropic = selectedProvider?.providerType === "anthropic";
- const isOpenAI = selectedProvider?.providerType === "openai";
- const anthropicTopEffort = effortLevels.includes("xhigh")
- ? "xhigh"
- : effortLevels.includes("high")
- ? "high"
- : clampedEffort;
- const openaiDefaultEffort = effortLevels.includes("high")
- ? "high"
- : effortLevels.includes("medium")
- ? "medium"
- : clampedEffort;
- const catalogDefaultEffort =
- reasoningCaps.defaultEffort &&
- effortLevels.includes(reasoningCaps.defaultEffort)
- ? reasoningCaps.defaultEffort
- : null;
- const nextReasoningEffort = reasoningCaps.supportsReasoning
- ? (catalogDefaultEffort ??
- (isAnthropic
- ? anthropicTopEffort
- : isOpenAI
- ? openaiDefaultEffort
- : effortLevels.includes("medium")
- ? "medium"
- : clampedEffort))
- : store.reasoningEffort;
+ const pinnedEffort = externalReasoningTakesEffort(reasoningCaps)
+ ? pinnedReasoningEffort(value, effortLevels)
+ : null;
+ const nextReasoningEffort = resolveExternalReasoningEffort({
+ caps: reasoningCaps,
+ providerType: selectedProvider?.providerType,
+ // The outgoing model's pin is what the live level holds, so an unpinned target resolves
+ // from the chat's own effort rather than inheriting one model's override.
+ current:
+ !pinnedEffort && pinHoldsLiveEffort()
+ ? (takeEffortDisplacedByPin() ?? store.reasoningEffort)
+ : store.reasoningEffort,
+ pinned: pinnedEffort,
+ });
+ // The chat's own level, before the pin takes its place, so clearing it can put it back.
+ if (pinnedEffort && nextReasoningEffort !== store.reasoningEffort) {
+ noteEffortDisplacedByPin(store.reasoningEffort);
+ }
// Clear any cached router-picked openrouter/free model unless staying on openrouter/free, else
// the chip keeps a stale ":<chosen>" suffix.
const stillOnOpenRouterFree =
@@ -3337,6 +3352,9 @@ export function ChatPage({
}
// Local model picked: drop any cached openrouter/free chosen model.
useChatRuntimeStore.setState({ lastOpenRouterChosenModel: null });
+ // The chat's own effort goes back where the load applies its own reasoning fields, since
+ // everything from here on can still abort or only queue a download, leaving the pinned
+ // model the one running.
void (async () => {
let showImageCompatibilityWarning = false;
if (view.mode === "single" && activeThreadId) {
diff --git a/studio/frontend/src/features/chat/chat-providers-dialog.tsx b/studio/frontend/src/features/chat/chat-providers-dialog.tsx
index 438c06c0709..7fcfefa6bd1 100644
--- a/studio/frontend/src/features/chat/chat-providers-dialog.tsx
+++ b/studio/frontend/src/features/chat/chat-providers-dialog.tsx
@@ -161,6 +161,11 @@ function formatModelSummary(models: string[]): string {
interface ChatProvidersSettingsProps {
providers: ExternalProviderConfig[];
onProvidersChange: (providers: ExternalProviderConfig[]) => void;
+ /** Open this connection's edit form on arrival instead of the list. Ignored until `providers`
+ * carries the id. */
+ openProviderId?: string | null;
+ /** Called once honoured, so the request cannot replay on a later visit. */
+ onOpenProviderConsumed?: () => void;
}
export function codexCapabilitiesWithPlanModels(
@@ -230,6 +235,8 @@ export function resolveCodexPickerModels(
export function ChatProvidersSettings({
providers,
onProvidersChange,
+ openProviderId = null,
+ onOpenProviderConsumed,
}: ChatProvidersSettingsProps) {
const providersRef = useRef(providers);
const seededProviderTypeRef = useRef<string | null>(null);
@@ -253,6 +260,7 @@ export function ChatProvidersSettings({
const [registry, setRegistry] = useState<ProviderRegistryEntry[]>([]);
const [availableModels, setAvailableModels] = useState<string[]>([]);
const [selectedModelIds, setSelectedModelIds] = useState<string[]>([]);
+ const [providersReady, setProvidersReady] = useState(false);
const [syncingProviders, setSyncingProviders] = useState(false);
const [registryLoading, setRegistryLoading] = useState(false);
const [modelsLoading, setModelsLoading] = useState(false);
@@ -444,6 +452,7 @@ export function ChatProvidersSettings({
// Trust the backend response. An empty array means every connection was removed, often from
// another tab; mirror that locally, else stale entries are un-removable here.
onProvidersChange(syncedProviders);
+ setProvidersReady(true);
// An empty list never says what this page is for, so open the form instead. Reads the synced
// response, not the local snapshot, so a stale empty list cannot flash the form at an
// existing user. Once only, else the focus re-sync would pull the user back here.
@@ -460,6 +469,9 @@ export function ChatProvidersSettings({
error instanceof Error ? error.message : "Unknown error";
toast.error(`Failed to load connections: ${message}`);
}
+ // A failed sync leaves the hydrated list as all there is. Waiting on a success the backend
+ // may never give would leave the deep link dead offline, with the gear opening nothing.
+ if (isMounted) setProvidersReady(true);
} finally {
if (isMounted && showSpinner) {
setRegistryLoading(false);
@@ -1206,6 +1218,26 @@ export function ChatProvidersSettings({
}
}
+ // Wait for backend data before opening a connection's form.
+ const openedProviderRef = useRef<string | null>(null);
+ useEffect(() => {
+ if (!openProviderId) {
+ openedProviderRef.current = null;
+ return;
+ }
+ if (!providersReady) return;
+ if (openedProviderRef.current === openProviderId) return;
+ const provider = providers.find(
+ (candidate) => candidate.id === openProviderId,
+ );
+ if (!provider) return;
+ openedProviderRef.current = openProviderId;
+ void editProvider(provider);
+ onOpenProviderConsumed?.();
+ // editProvider is redeclared each render; the latch above is what fires this once per id.
+ // eslint-disable-next-line react-hooks/exhaustive-deps
+ }, [openProviderId, providers, providersReady, onOpenProviderConsumed]);
+
async function deleteProvider(providerId: string) {
setMutatingProvider(true);
try {
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
index f746eab20a1..3b79aaea1da 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
@@ -80,12 +80,14 @@ import {
loadedGpuMemoryFields,
noteLoadedModelReasoningMode,
persistGpuMemoryModeOnLoad,
+ pinHoldsLiveEffort,
readPersistedGpuMemoryMode,
readPersistedSpeculativeType,
reconcilePersistedGpuIds,
resolvePreserveThinkingOnLoad,
resolveToolsEnabledOnLoad,
saveSpeculativeType,
+ takeEffortDisplacedByPin,
useChatRuntimeStore,
type LoadingModelPick,
type ReasoningEffort,
@@ -1965,7 +1967,12 @@ export function useChatModelRuntime() {
loadResponse.reasoning_effort_levels.length > 0
? (loadResponse.reasoning_effort_levels as ReasoningEffort[])
: (["low", "medium", "high"] as const);
- const existingReasoningEffort = useChatRuntimeStore.getState().reasoningEffort;
+ // The chat's own level when the model this replaces was running a pin's, for the
+ // reason applyActiveModelStatusToStore gives at its own clamp: a pin is one model's,
+ // and everything between the pick and this response can abort without loading.
+ const existingReasoningEffort =
+ (pinHoldsLiveEffort() ? takeEffortDisplacedByPin() : null) ??
+ useChatRuntimeStore.getState().reasoningEffort;
const clampedReasoningEffort =
reasoningStyle === "enable_thinking_effort" ||
reasoningStyle === "reasoning_effort"
diff --git a/studio/frontend/src/features/chat/index.ts b/studio/frontend/src/features/chat/index.ts
index 96f198162bb..5cea8f5cffd 100644
--- a/studio/frontend/src/features/chat/index.ts
+++ b/studio/frontend/src/features/chat/index.ts
@@ -201,6 +201,14 @@ export {
isExternalModelId,
parseExternalModelId,
} from "./external-providers";
+// A provider catalogue lands async, so capability reads need to re-run when it does.
+export { modelCatalogVersion, subscribeModelCatalog } from "./model-catalog";
+// What a per-model reasoning pin displaced in the live runtime, so clearing it can put it back.
+export {
+ noteEffortDisplacedByPin,
+ reconcilePinnedReasoningEffort,
+ takeEffortDisplacedByPin,
+} from "./stores/chat-runtime-store";
export {
type AttachmentText,
assertDocumentAttachmentSize,
diff --git a/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts b/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
index f79c0386732..a24b163badf 100644
--- a/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
+++ b/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
@@ -25,8 +25,10 @@ import {
loadedGpuMemoryFields,
normalizeSpeculativeType,
noteLoadedModelReasoningMode,
+ pinHoldsLiveEffort,
resolvePreserveThinkingOnLoad,
resolveToolsEnabledOnLoad,
+ takeEffortDisplacedByPin,
useChatRuntimeStore,
} from "../stores/chat-runtime-store";
import {
@@ -205,14 +207,18 @@ export function applyActiveModelStatusToStore(
const storedReasoningEnabled = loadOptionalBool(CHAT_REASONING_ENABLED_KEY);
const currentSpecType = normalizeSpeculativeType(status.speculative_type);
const prevState = useChatRuntimeStore.getState();
+ // The chat's own level when the model this replaces was running a per-model pin's: the clamp
+ // narrows the chat's level to this model's ladder, and a pin is one model's, not the chat's.
+ // Taken here rather than where the model was picked, because this is a model that has actually
+ // become resident: everything from the pick to here can still abort, or only queue a download.
+ const effortToClamp =
+ (pinHoldsLiveEffort() ? takeEffortDisplacedByPin() : null) ??
+ prevState.reasoningEffort;
const clampedReasoningEffort =
reasoningStyle === "enable_thinking_effort" ||
reasoningStyle === "reasoning_effort"
- ? clampReasoningEffortToLevels(
- prevState.reasoningEffort,
- reasoningEffortLevels,
- )
- : clampLocalReasoningEffort(prevState.reasoningEffort);
+ ? clampReasoningEffortToLevels(effortToClamp, reasoningEffortLevels)
+ : clampLocalReasoningEffort(effortToClamp);
const nextDefaultChatTemplate =
status.chat_template === undefined
? prevState.defaultChatTemplate
diff --git a/studio/frontend/src/features/chat/model-catalog-snapshot.ts b/studio/frontend/src/features/chat/model-catalog-snapshot.ts
index fd43ba66475..4a3198de333 100644
--- a/studio/frontend/src/features/chat/model-catalog-snapshot.ts
+++ b/studio/frontend/src/features/chat/model-catalog-snapshot.ts
@@ -1,13 +1,15 @@
// SPDX-License-Identifier: AGPL-3.0-only
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-// Generated by scripts/refresh-model-catalog-snapshot.mjs from https://models.dev/api.json on 2026-09-14. Do not edit by hand.
+// Generated by scripts/refresh-model-catalog-snapshot.mjs from https://models.dev/api.json on 2026-09-17. Do not edit by hand.
export interface ModelCatalogSnapshotEntry {
reasoning?: true;
efforts?: readonly string[];
toggle?: true;
input?: readonly string[];
+ /** Published context window in tokens. */
+ context?: number;
}
export const MODEL_CATALOG_SNAPSHOT: Readonly<
@@ -15,709 +17,712 @@ export const MODEL_CATALOG_SNAPSHOT: Readonly<
> = {
"openai": {
"chatgpt-image-latest": {"input":["text","image"]},
- "gpt-3.5-turbo": {"input":["text"]},
- "gpt-4": {"input":["text"]},
- "gpt-4-turbo": {"input":["text","image"]},
- "gpt-4.1": {"input":["text","image","pdf"]},
- "gpt-4.1-mini": {"input":["text","image","pdf"]},
- "gpt-4.1-nano": {"input":["text","image"]},
- "gpt-4o": {"input":["text","image","pdf"]},
- "gpt-4o-2024-05-13": {"input":["text","image"]},
- "gpt-4o-2024-08-06": {"input":["text","image"]},
- "gpt-4o-2024-11-20": {"input":["text","image"]},
- "gpt-4o-mini": {"input":["text","image","pdf"]},
- "gpt-5": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image"]},
- "gpt-5-mini": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image"]},
- "gpt-5-nano": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image"]},
- "gpt-5-pro": {"reasoning":true,"efforts":["high"],"input":["text","image"]},
- "gpt-5.1": {"reasoning":true,"efforts":["none","low","medium","high"],"input":["text","image"]},
- "gpt-5.2": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image"]},
- "gpt-5.2-chat-latest": {"reasoning":true,"efforts":["medium"],"input":["text","image"]},
- "gpt-5.2-pro": {"reasoning":true,"efforts":["medium","high","xhigh"],"input":["text","image"]},
- "gpt-5.3-chat-latest": {"input":["text","image"]},
- "gpt-5.3-codex": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image","pdf"]},
- "gpt-5.3-codex-spark": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image","pdf"]},
- "gpt-5.4": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image","pdf"]},
- "gpt-5.4-mini": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image"]},
- "gpt-5.4-nano": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image"]},
- "gpt-5.4-pro": {"reasoning":true,"efforts":["medium","high","xhigh"],"input":["text","image"]},
- "gpt-5.5": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image","pdf"]},
- "gpt-5.5-pro": {"reasoning":true,"efforts":["medium","high","xhigh"],"input":["text","image","pdf"]},
- "gpt-5.6": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "gpt-5.6-luna": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "gpt-5.6-sol": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "gpt-5.6-terra": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "gpt-6-astra": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
+ "gpt-3.5-turbo": {"input":["text"],"context":16385},
+ "gpt-4": {"input":["text"],"context":8192},
+ "gpt-4-turbo": {"input":["text","image"],"context":128000},
+ "gpt-4.1": {"input":["text","image","pdf"],"context":1047576},
+ "gpt-4.1-mini": {"input":["text","image","pdf"],"context":1047576},
+ "gpt-4.1-nano": {"input":["text","image"],"context":1047576},
+ "gpt-4o": {"input":["text","image","pdf"],"context":128000},
+ "gpt-4o-2024-05-13": {"input":["text","image"],"context":128000},
+ "gpt-4o-2024-08-06": {"input":["text","image"],"context":128000},
+ "gpt-4o-2024-11-20": {"input":["text","image"],"context":128000},
+ "gpt-4o-mini": {"input":["text","image","pdf"],"context":128000},
+ "gpt-5": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image"],"context":400000},
+ "gpt-5-mini": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image"],"context":400000},
+ "gpt-5-nano": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image"],"context":400000},
+ "gpt-5-pro": {"reasoning":true,"efforts":["high"],"input":["text","image"],"context":400000},
+ "gpt-5.1": {"reasoning":true,"efforts":["none","low","medium","high"],"input":["text","image"],"context":400000},
+ "gpt-5.2": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image"],"context":400000},
+ "gpt-5.2-chat-latest": {"reasoning":true,"efforts":["medium"],"input":["text","image"],"context":128000},
+ "gpt-5.2-pro": {"reasoning":true,"efforts":["medium","high","xhigh"],"input":["text","image"],"context":400000},
+ "gpt-5.3-chat-latest": {"input":["text","image"],"context":128000},
+ "gpt-5.3-codex": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image","pdf"],"context":400000},
+ "gpt-5.3-codex-spark": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image","pdf"],"context":128000},
+ "gpt-5.4": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image","pdf"],"context":1050000},
+ "gpt-5.4-mini": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image"],"context":400000},
+ "gpt-5.4-nano": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image"],"context":400000},
+ "gpt-5.4-pro": {"reasoning":true,"efforts":["medium","high","xhigh"],"input":["text","image"],"context":1050000},
+ "gpt-5.5": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image","pdf"],"context":1050000},
+ "gpt-5.5-pro": {"reasoning":true,"efforts":["medium","high","xhigh"],"input":["text","image","pdf"],"context":1050000},
+ "gpt-5.6": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1050000},
+ "gpt-5.6-luna": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1050000},
+ "gpt-5.6-sol": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1050000},
+ "gpt-5.6-terra": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1050000},
+ "gpt-6-astra": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1050000},
"gpt-image-1": {"input":["text","image"]},
"gpt-image-1-mini": {"input":["text","image"]},
"gpt-image-1.5": {"input":["text","image"]},
"gpt-image-2": {"input":["text","image"]},
- "gpt-realtime-2.1": {"reasoning":true,"efforts":["minimal","low","medium","high","xhigh"],"input":["text","audio","image"]},
- "o1": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","pdf"]},
- "o1-pro": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image"]},
- "o3": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","pdf"]},
- "o3-mini": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"]},
- "o3-pro": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image"]},
- "o4-mini": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image"]},
- "text-embedding-3-large": {"input":["text"]},
- "text-embedding-3-small": {"input":["text"]},
- "text-embedding-ada-002": {"input":["text"]},
+ "gpt-realtime-2.1": {"reasoning":true,"efforts":["minimal","low","medium","high","xhigh"],"input":["text","audio","image"],"context":128000},
+ "o1": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","pdf"],"context":200000},
+ "o1-pro": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image"],"context":200000},
+ "o3": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","pdf"],"context":200000},
+ "o3-mini": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"],"context":200000},
+ "o3-pro": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image"],"context":200000},
+ "o4-mini": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image"],"context":200000},
+ "text-embedding-3-large": {"input":["text"],"context":8191},
+ "text-embedding-3-small": {"input":["text"],"context":8191},
+ "text-embedding-ada-002": {"input":["text"],"context":8192},
},
"anthropic": {
- "claude-fable-5": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "claude-fable-5-1": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "claude-haiku-4-5": {"reasoning":true,"input":["text","image","pdf"]},
- "claude-haiku-4-5-20251001": {"reasoning":true,"input":["text","image","pdf"]},
- "claude-opus-4-5": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","pdf"]},
- "claude-opus-4-5-20251101": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","pdf"]},
- "claude-opus-4-6": {"reasoning":true,"efforts":["low","medium","high","max"],"input":["text","image","pdf"]},
- "claude-opus-4-7": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "claude-opus-4-8": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "claude-opus-5": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "claude-sonnet-4-5": {"reasoning":true,"input":["text","image","pdf"]},
- "claude-sonnet-4-5-20250929": {"reasoning":true,"input":["text","image","pdf"]},
- "claude-sonnet-4-6": {"reasoning":true,"efforts":["low","medium","high","max"],"input":["text","image","pdf"]},
- "claude-sonnet-5": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"toggle":true,"input":["text","image","pdf"]},
+ "claude-fable-5": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1000000},
+ "claude-fable-5-1": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1000000},
+ "claude-haiku-4-5": {"reasoning":true,"input":["text","image","pdf"],"context":200000},
+ "claude-haiku-4-5-20251001": {"reasoning":true,"input":["text","image","pdf"],"context":200000},
+ "claude-opus-4-5": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","pdf"],"context":200000},
+ "claude-opus-4-5-20251101": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","pdf"],"context":200000},
+ "claude-opus-4-6": {"reasoning":true,"efforts":["low","medium","high","max"],"input":["text","image","pdf"],"context":1000000},
+ "claude-opus-4-7": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1000000},
+ "claude-opus-4-8": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1000000},
+ "claude-opus-5": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1000000},
+ "claude-sonnet-4-5": {"reasoning":true,"input":["text","image","pdf"],"context":1000000},
+ "claude-sonnet-4-5-20250929": {"reasoning":true,"input":["text","image","pdf"],"context":1000000},
+ "claude-sonnet-4-6": {"reasoning":true,"efforts":["low","medium","high","max"],"input":["text","image","pdf"],"context":1000000},
+ "claude-sonnet-5": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"toggle":true,"input":["text","image","pdf"],"context":1000000},
},
"openrouter": {
- "~anthropic/claude-fable-latest": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "~anthropic/claude-haiku-latest": {"reasoning":true,"toggle":true,"input":["text","image","pdf"]},
- "~anthropic/claude-opus-latest": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"toggle":true,"input":["text","image","pdf"]},
- "~anthropic/claude-sonnet-latest": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"toggle":true,"input":["text","image","pdf"]},
- "~deepseek/deepseek-v4-flash-latest": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text"]},
- "~google/gemini-flash-latest": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","pdf","audio"]},
- "~google/gemini-pro-latest": {"reasoning":true,"efforts":["low","medium","high"],"input":["audio","pdf","image","text","video"]},
- "~moonshotai/kimi-latest": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image","video"]},
- "~openai/gpt-astra-latest": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["pdf","image","text"]},
- "~openai/gpt-luna-latest": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["pdf","image","text"]},
- "~openai/gpt-mini-latest": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["pdf","image","text"]},
- "~openai/gpt-sol-latest": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["pdf","image","text"]},
- "~openai/gpt-terra-latest": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["pdf","image","text"]},
- "~x-ai/grok-latest": {"reasoning":true,"efforts":["low","medium","high","xhigh"],"input":["text","image","pdf"]},
- "~z-ai/glm-flash-latest": {"reasoning":true,"efforts":["low","high","max"],"input":["text","image","video"]},
- "~z-ai/glm-latest": {"reasoning":true,"efforts":["low","high","max"],"input":["text"]},
- "aion-labs/aion-2.0": {"reasoning":true,"input":["text"]},
- "aion-labs/aion-3.0": {"reasoning":true,"input":["text"]},
- "aion-labs/aion-3.0-mini": {"reasoning":true,"input":["text"]},
- "aion-labs/aion-rp-llama-3.1-8b": {"input":["text"]},
- "amazon/nova-2-lite-v1": {"reasoning":true,"toggle":true,"input":["text","image","video","pdf"]},
- "amazon/nova-lite-v1": {"input":["text","image"]},
- "amazon/nova-micro-v1": {"input":["text"]},
- "amazon/nova-premier-v1": {"input":["text","image"]},
- "amazon/nova-pro-v1": {"input":["text","image"]},
- "anthracite-org/magnum-v4-72b": {"input":["text"]},
- "anthropic/claude-3-haiku": {"input":["text","image"]},
- "anthropic/claude-fable-5": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "anthropic/claude-fable-5.1": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "anthropic/claude-haiku-4.5": {"reasoning":true,"toggle":true,"input":["text","image","pdf"]},
- "anthropic/claude-opus-4": {"reasoning":true,"toggle":true,"input":["image","text","pdf"]},
- "anthropic/claude-opus-4.1": {"reasoning":true,"toggle":true,"input":["text","image","pdf"]},
- "anthropic/claude-opus-4.5": {"reasoning":true,"toggle":true,"input":["text","image","pdf"]},
- "anthropic/claude-opus-4.6": {"reasoning":true,"efforts":["low","medium","high","max"],"toggle":true,"input":["text","image","pdf"]},
- "anthropic/claude-opus-4.7": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"toggle":true,"input":["text","image","pdf"]},
- "anthropic/claude-opus-4.8": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"toggle":true,"input":["text","image","pdf"]},
- "anthropic/claude-opus-5": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"toggle":true,"input":["text","image","pdf"]},
- "anthropic/claude-sonnet-4": {"reasoning":true,"toggle":true,"input":["image","text","pdf"]},
- "anthropic/claude-sonnet-4.5": {"reasoning":true,"toggle":true,"input":["text","image","pdf"]},
- "anthropic/claude-sonnet-4.6": {"reasoning":true,"efforts":["low","medium","high","max"],"toggle":true,"input":["text","image","pdf"]},
- "anthropic/claude-sonnet-5": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"toggle":true,"input":["text","image","pdf"]},
- "arcee-ai/trinity-large-thinking": {"reasoning":true,"input":["text"]},
- "baidu/ernie-4.5-vl-424b-a47b": {"reasoning":true,"toggle":true,"input":["image","text"]},
- "bytedance-seed/seed-1.6": {"reasoning":true,"toggle":true,"input":["image","text","video"]},
- "bytedance-seed/seed-1.6-flash": {"reasoning":true,"toggle":true,"input":["image","text","video"]},
- "bytedance-seed/seed-2-1-turbo": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "bytedance-seed/seed-2.0-code": {"reasoning":true,"efforts":["low","medium","high"],"toggle":true,"input":["text","image","video"]},
- "bytedance-seed/seed-2.0-lite": {"reasoning":true,"efforts":["minimal","low","medium","high"],"toggle":true,"input":["text","image","video"]},
- "bytedance-seed/seed-2.0-mini": {"reasoning":true,"efforts":["minimal","low","medium","high"],"toggle":true,"input":["text","image","video"]},
- "bytedance/ui-tars-1.5-7b": {"input":["image","text"]},
- "cognitivecomputations/dolphin-mistral-24b-venice-edition": {"input":["text"]},
- "cohere/command-a": {"input":["text"]},
- "cohere/command-r-08-2024": {"input":["text"]},
- "cohere/command-r-plus-08-2024": {"input":["text"]},
- "cohere/command-r7b-12-2024": {"input":["text"]},
- "cohere/north-mini-code:free": {"reasoning":true,"toggle":true,"input":["text"]},
- "deepseek/deepseek-chat": {"input":["text"]},
- "deepseek/deepseek-chat-v3-0324": {"input":["text"]},
- "deepseek/deepseek-chat-v3.1": {"reasoning":true,"toggle":true,"input":["text"]},
- "deepseek/deepseek-r1": {"reasoning":true,"input":["text"]},
- "deepseek/deepseek-r1-0528": {"reasoning":true,"input":["text"]},
- "deepseek/deepseek-r1-distill-llama-70b": {"reasoning":true,"toggle":true,"input":["text"]},
- "deepseek/deepseek-v3.1-terminus": {"reasoning":true,"toggle":true,"input":["text"]},
- "deepseek/deepseek-v3.2": {"reasoning":true,"toggle":true,"input":["text"]},
- "deepseek/deepseek-v3.2-exp": {"reasoning":true,"toggle":true,"input":["text"]},
- "deepseek/deepseek-v4-flash": {"reasoning":true,"efforts":["high","xhigh"],"toggle":true,"input":["text"]},
- "deepseek/deepseek-v4-flash-0731": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text"]},
- "deepseek/deepseek-v4-flash-vision-exp": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image"]},
- "deepseek/deepseek-v4-pro": {"reasoning":true,"efforts":["high","xhigh"],"toggle":true,"input":["text"]},
- "deepseek/deepseek-v4-pro-0813": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text"]},
- "deepseek/deepseek-v4.1-flash": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image"]},
- "dots-studio/dots-3-note-preview:free": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "google/gemini-2.5-flash": {"reasoning":true,"toggle":true,"input":["text","image","audio","video","pdf"]},
- "google/gemini-2.5-flash-image": {"input":["text","image"]},
- "google/gemini-2.5-flash-lite": {"reasoning":true,"toggle":true,"input":["text","image","audio","video","pdf"]},
- "google/gemini-2.5-pro": {"reasoning":true,"input":["text","image","audio","video","pdf"]},
- "google/gemini-2.5-pro-preview": {"reasoning":true,"input":["pdf","image","text","audio"]},
- "google/gemini-2.5-pro-preview-05-06": {"reasoning":true,"input":["text","image","pdf","audio","video"]},
- "google/gemini-3-flash-preview": {"reasoning":true,"efforts":["minimal","low","medium","high"],"toggle":true,"input":["text","image","video","audio","pdf"]},
- "google/gemini-3-pro-image": {"reasoning":true,"input":["text","image"]},
- "google/gemini-3-pro-image-preview": {"reasoning":true,"input":["text","image"]},
- "google/gemini-3.1-flash-image": {"reasoning":true,"efforts":["minimal","high"],"toggle":true,"input":["image","text"]},
- "google/gemini-3.1-flash-image-preview": {"reasoning":true,"efforts":["minimal","high"],"toggle":true,"input":["image","text"]},
- "google/gemini-3.1-flash-lite": {"reasoning":true,"efforts":["minimal","low","medium","high"],"toggle":true,"input":["text","image","video","audio","pdf"]},
- "google/gemini-3.1-flash-lite-image": {"reasoning":true,"efforts":["minimal","high"],"toggle":true,"input":["text","image"]},
- "google/gemini-3.1-flash-lite-preview": {"reasoning":true,"efforts":["minimal","low","medium","high"],"toggle":true,"input":["text","image","video","audio","pdf"]},
- "google/gemini-3.1-pro-preview": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "google/gemini-3.1-pro-preview-customtools": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "google/gemini-3.5-flash": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "google/gemini-3.5-flash-lite": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "google/gemini-3.6-flash": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "google/gemini-3.7-flash": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "google/gemini-3.8-flash": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "google/gemma-2-27b-it": {"input":["text"]},
- "google/gemma-3-12b-it": {"input":["text","image"]},
- "google/gemma-3-27b-it": {"input":["text","image"]},
- "google/gemma-3-4b-it": {"input":["text","image"]},
- "google/gemma-4-26b-a4b-it": {"reasoning":true,"toggle":true,"input":["image","text","video"]},
- "google/gemma-4-26b-a4b-it:free": {"reasoning":true,"toggle":true,"input":["image","text","video"]},
- "google/gemma-4-31b-it": {"reasoning":true,"toggle":true,"input":["image","text","video"]},
- "google/gemma-4-31b-it:free": {"reasoning":true,"toggle":true,"input":["image","text","video"]},
- "google/lyria-3-clip-preview": {"input":["text","image"]},
- "google/lyria-3-pro-preview": {"input":["text","image"]},
- "gryphe/mythomax-l2-13b": {"input":["text"]},
- "ibm-granite/granite-4.0-h-micro": {"input":["text"]},
- "ibm-granite/granite-4.2-8b": {"reasoning":true,"efforts":["none","low","high"],"input":["text"]},
- "inception/mercury-2": {"reasoning":true,"efforts":["none","low","medium","high"],"input":["text"]},
- "inception/mercury-2.5": {"reasoning":true,"efforts":["none","low","medium","high"],"input":["text"]},
- "inclusionai/ling-3.0-flash": {"reasoning":true,"toggle":true,"input":["text"]},
- "inclusionai/ling-3.0-flash-fin": {"reasoning":true,"toggle":true,"input":["text"]},
- "inclusionai/ling-3.0-flash-fin:free": {"reasoning":true,"toggle":true,"input":["text"]},
- "inclusionai/ling-3.0-flash-sante:free": {"reasoning":true,"toggle":true,"input":["text"]},
- "inclusionai/ling-3.0-flash-vl": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "inclusionai/ling-3.0-flash-vl:free": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "inference-net/schematron-v2-small": {"input":["text"]},
- "inference-net/schematron-v2-turbo": {"input":["text"]},
- "kwaipilot/kat-coder-pro-v2": {"input":["text"]},
- "kwaipilot/kat-coder-pro-v2.5": {"input":["text"]},
- "liquid/lfm-2.5-2.6b:free": {"reasoning":true,"input":["text"]},
- "mancer/weaver": {"input":["text"]},
- "meituan/longcat-2.0": {"reasoning":true,"toggle":true,"input":["text"]},
- "meta-llama/llama-3.1-70b-instruct": {"input":["text"]},
- "meta-llama/llama-3.1-8b-instruct": {"input":["text"]},
- "meta-llama/llama-3.2-1b-instruct": {"input":["text"]},
- "meta-llama/llama-3.2-3b-instruct": {"input":["text"]},
- "meta-llama/llama-3.3-70b-instruct": {"input":["text"]},
- "meta-llama/llama-4-maverick": {"input":["text","image"]},
- "meta-llama/llama-4-scout": {"input":["text","image"]},
- "meta-llama/llama-guard-4-12b": {"input":["image","text"]},
- "meta/muse-glimmer-30b": {"reasoning":true,"efforts":["low","medium","high","xhigh"],"input":["text","image"]},
- "meta/muse-spark-1.1": {"reasoning":true,"efforts":["minimal","low","medium","high","xhigh"],"input":["text","image","video","pdf","audio"]},
- "meta/muse-spark-1.2": {"reasoning":true,"efforts":["minimal","low","medium","high","xhigh"],"input":["text","image","video","pdf","audio"]},
- "meta/muse-spark-1.2-contributor": {"reasoning":true,"efforts":["minimal","low","medium","high","xhigh"],"input":["text","image","video","pdf","audio"]},
- "meta/muse-spark-1.3": {"reasoning":true,"efforts":["minimal","low","medium","high","xhigh","max"],"input":["text","image","video","pdf","audio"]},
- "meta/muse-spark-1.3-contributor": {"reasoning":true,"efforts":["minimal","low","medium","high","xhigh","max"],"input":["text","image","video","pdf","audio"]},
- "microsoft/phi-4": {"input":["text"]},
- "microsoft/wizardlm-2-8x22b": {"input":["text"]},
- "minimax/minimax-01": {"input":["text","image"]},
- "minimax/minimax-m1": {"reasoning":true,"toggle":true,"input":["text"]},
- "minimax/minimax-m2": {"reasoning":true,"input":["text"]},
- "minimax/minimax-m2-her": {"input":["text"]},
- "minimax/minimax-m2.1": {"reasoning":true,"input":["text"]},
- "minimax/minimax-m2.5": {"reasoning":true,"input":["text"]},
- "minimax/minimax-m2.7": {"reasoning":true,"input":["text"]},
- "minimax/minimax-m3": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "mistralai/codestral-2508": {"input":["text","pdf"]},
- "mistralai/devstral-2512": {"input":["text","pdf"]},
- "mistralai/ministral-14b-2512": {"input":["text","image"]},
- "mistralai/ministral-3b-2512": {"input":["text","image"]},
- "mistralai/ministral-8b-2512": {"input":["text","image"]},
- "mistralai/mistral-large": {"input":["text","pdf"]},
- "mistralai/mistral-large-2407": {"input":["text","pdf"]},
- "mistralai/mistral-large-2512": {"input":["text","image","pdf"]},
- "mistralai/mistral-medium-3": {"input":["text","image","pdf"]},
- "mistralai/mistral-medium-3-5": {"reasoning":true,"efforts":["none","high"],"input":["text","image","pdf"]},
- "mistralai/mistral-medium-3.1": {"input":["text","image","pdf"]},
- "mistralai/mistral-nemo": {"input":["text"]},
- "mistralai/mistral-saba": {"input":["text","pdf"]},
- "mistralai/mistral-small-24b-instruct-2501": {"input":["text"]},
- "mistralai/mistral-small-2603": {"reasoning":true,"efforts":["none","high"],"input":["text","image"]},
- "mistralai/mistral-small-3.1-24b-instruct": {"input":["text","image"]},
- "mistralai/mistral-small-3.2-24b-instruct": {"input":["image","text"]},
- "mistralai/mixtral-8x22b-instruct": {"input":["text","pdf"]},
- "mistralai/voxtral-small-24b-2507": {"input":["text","audio","pdf"]},
- "moonshotai/kimi-k2": {"input":["text"]},
- "moonshotai/kimi-k2-0905": {"input":["text"]},
- "moonshotai/kimi-k2-thinking": {"reasoning":true,"input":["text"]},
- "moonshotai/kimi-k2.5": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "moonshotai/kimi-k2.6": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "moonshotai/kimi-k2.7-code": {"reasoning":true,"input":["text","image"]},
- "moonshotai/kimi-k3": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image","video"]},
- "morph/morph-v3-fast": {"input":["text"]},
- "morph/morph-v3-large": {"input":["text"]},
- "nex-agi/nex-n2.5-mini:free": {"reasoning":true,"efforts":["none","medium","high"],"input":["text","image"]},
- "nex-agi/nex-n2.5-pro:free": {"reasoning":true,"efforts":["none","medium","high"],"input":["text","image"]},
- "nousresearch/hermes-3-llama-3.1-405b": {"input":["text"]},
- "nousresearch/hermes-3-llama-3.1-70b": {"input":["text"]},
- "nousresearch/hermes-4-405b": {"reasoning":true,"toggle":true,"input":["text"]},
- "nvidia/nemotron-3-nano-30b-a3b": {"reasoning":true,"toggle":true,"input":["text"]},
- "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free": {"reasoning":true,"toggle":true,"input":["text","image","video","audio"]},
- "nvidia/nemotron-3-super-120b-a12b": {"reasoning":true,"efforts":["low","medium"],"toggle":true,"input":["text"]},
- "nvidia/nemotron-3-super-120b-a12b:free": {"reasoning":true,"efforts":["low","medium"],"toggle":true,"input":["text"]},
- "nvidia/nemotron-3-ultra-550b-a55b": {"reasoning":true,"efforts":["medium","high"],"toggle":true,"input":["text"]},
- "nvidia/nemotron-3-ultra-550b-a55b:free": {"reasoning":true,"efforts":["medium","high"],"toggle":true,"input":["text"]},
- "nvidia/nemotron-3.5-content-safety": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "nvidia/nemotron-3.5-content-safety:free": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "nvidia/nemotron-3.5-lightning": {"reasoning":true,"toggle":true,"input":["text"]},
- "nvidia/nemotron-3.5-lightning:free": {"reasoning":true,"toggle":true,"input":["text"]},
- "openai/gpt-3.5-turbo": {"input":["text"]},
- "openai/gpt-3.5-turbo-0613": {"input":["text"]},
- "openai/gpt-3.5-turbo-16k": {"input":["text"]},
- "openai/gpt-3.5-turbo-instruct": {"input":["text"]},
- "openai/gpt-4": {"input":["text"]},
- "openai/gpt-4-turbo": {"input":["text","image"]},
- "openai/gpt-4-turbo-preview": {"input":["text"]},
- "openai/gpt-4.1": {"input":["text","image","pdf"]},
- "openai/gpt-4.1-mini": {"input":["text","image","pdf"]},
- "openai/gpt-4.1-nano": {"input":["image","text","pdf"]},
- "openai/gpt-4o": {"input":["text","image","pdf"]},
- "openai/gpt-4o-2024-05-13": {"input":["text","image","pdf"]},
- "openai/gpt-4o-2024-08-06": {"input":["text","image","pdf"]},
- "openai/gpt-4o-2024-11-20": {"input":["text","image","pdf"]},
- "openai/gpt-4o-mini": {"input":["text","image","pdf"]},
- "openai/gpt-4o-mini-2024-07-18": {"input":["text","image","pdf"]},
- "openai/gpt-5": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","pdf"]},
- "openai/gpt-5-image": {"reasoning":true,"input":["image","text","pdf"]},
- "openai/gpt-5-image-mini": {"reasoning":true,"input":["pdf","image","text"]},
- "openai/gpt-5-mini": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","pdf"]},
- "openai/gpt-5-nano": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","pdf"]},
- "openai/gpt-5-pro": {"reasoning":true,"efforts":["high"],"input":["image","text","pdf"]},
- "openai/gpt-5.1": {"reasoning":true,"efforts":["none","low","medium","high"],"input":["image","text","pdf"]},
- "openai/gpt-5.1-codex": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image"]},
- "openai/gpt-5.1-codex-max": {"reasoning":true,"efforts":["low","medium","high","xhigh"],"input":["text","image"]},
- "openai/gpt-5.1-codex-mini": {"reasoning":true,"efforts":["low","medium","high"],"toggle":true,"input":["text","image"]},
- "openai/gpt-5.2": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["pdf","image","text"]},
- "openai/gpt-5.2-chat": {"input":["pdf","image","text"]},
- "openai/gpt-5.2-codex": {"reasoning":true,"efforts":["low","medium","high","xhigh"],"input":["text","image"]},
- "openai/gpt-5.2-pro": {"reasoning":true,"efforts":["medium","high","xhigh"],"input":["image","text","pdf"]},
- "openai/gpt-5.3-codex": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image","pdf"]},
- "openai/gpt-5.4": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image","pdf"]},
- "openai/gpt-5.4-image-2": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["image","text","pdf"]},
- "openai/gpt-5.4-mini": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["pdf","image","text"]},
- "openai/gpt-5.4-nano": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["pdf","image","text"]},
- "openai/gpt-5.4-pro": {"reasoning":true,"efforts":["medium","high","xhigh"],"input":["text","image","pdf"]},
- "openai/gpt-5.5": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image","pdf"]},
- "openai/gpt-5.5-pro": {"reasoning":true,"efforts":["medium","high","xhigh"],"input":["text","image","pdf"]},
- "openai/gpt-5.6-luna": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "openai/gpt-5.6-luna-pro": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "openai/gpt-5.6-sol": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "openai/gpt-5.6-sol-pro": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "openai/gpt-5.6-terra": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "openai/gpt-5.6-terra-pro": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "openai/gpt-6-astra": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"]},
- "openai/gpt-6-astra-pro": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["pdf","image","text"]},
- "openai/gpt-audio": {"input":["text","audio"]},
- "openai/gpt-audio-mini": {"input":["text","audio"]},
- "openai/gpt-chat-latest": {"input":["text","image","pdf"]},
- "openai/gpt-oss-120b": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"]},
- "openai/gpt-oss-20b": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"]},
- "openai/gpt-oss-safeguard-20b": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"]},
- "openai/o1": {"reasoning":true,"toggle":true,"input":["text","image","pdf"]},
- "openai/o1-pro": {"reasoning":true,"toggle":true,"input":["text","image","pdf"]},
- "openai/o3": {"reasoning":true,"toggle":true,"input":["text","image","pdf"]},
- "openai/o3-mini": {"reasoning":true,"toggle":true,"input":["text","pdf"]},
- "openai/o3-mini-high": {"reasoning":true,"efforts":["high"],"input":["text","pdf"]},
- "openai/o3-pro": {"reasoning":true,"toggle":true,"input":["text","pdf","image"]},
- "openai/o4-mini": {"reasoning":true,"toggle":true,"input":["image","text","pdf"]},
- "openai/o4-mini-high": {"reasoning":true,"efforts":["high"],"input":["image","text","pdf"]},
- "openrouter/auto": {"reasoning":true,"input":["text","image","audio","pdf","video"]},
- "openrouter/bodybuilder": {"input":["text"]},
- "openrouter/free": {"reasoning":true,"input":["text","image"]},
- "openrouter/fusion": {"input":["text"]},
- "openrouter/pareto-code": {"input":["text"]},
- "perceptron/perceptron-mk1": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "perplexity/sonar": {"input":["text","image"]},
- "perplexity/sonar-deep-research": {"reasoning":true,"toggle":true,"input":["text"]},
- "perplexity/sonar-pro": {"input":["text","image"]},
- "perplexity/sonar-pro-search": {"reasoning":true,"input":["text","image"]},
- "perplexity/sonar-reasoning-pro": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "poolside/laguna-s-2.1": {"reasoning":true,"toggle":true,"input":["text"]},
- "poolside/laguna-s-2.1:free": {"reasoning":true,"toggle":true,"input":["text"]},
- "poolside/laguna-xs-2.1": {"reasoning":true,"toggle":true,"input":["text"]},
- "poolside/laguna-xs-2.1:free": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen/qwen-2.5-72b-instruct": {"input":["text"]},
- "qwen/qwen-2.5-7b-instruct": {"input":["text"]},
- "qwen/qwen-2.5-coder-32b-instruct": {"input":["text"]},
- "qwen/qwen-plus": {"input":["text"]},
- "qwen/qwen-plus-2025-07-28": {"input":["text"]},
- "qwen/qwen2.5-vl-72b-instruct": {"input":["text","image"]},
- "qwen/qwen3-14b": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen/qwen3-235b-a22b": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen/qwen3-235b-a22b-2507": {"input":["text"]},
- "qwen/qwen3-235b-a22b-thinking-2507": {"reasoning":true,"input":["text"]},
- "qwen/qwen3-30b-a3b": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen/qwen3-30b-a3b-instruct-2507": {"input":["text"]},
- "qwen/qwen3-30b-a3b-thinking-2507": {"reasoning":true,"input":["text"]},
- "qwen/qwen3-32b": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen/qwen3-8b": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen/qwen3-coder": {"input":["text"]},
- "qwen/qwen3-coder-30b-a3b-instruct": {"input":["text"]},
- "qwen/qwen3-coder-flash": {"input":["text"]},
- "qwen/qwen3-coder-next": {"input":["text"]},
- "qwen/qwen3-coder-plus": {"input":["text"]},
- "qwen/qwen3-max": {"input":["text"]},
- "qwen/qwen3-max-thinking": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen/qwen3-next-80b-a3b-instruct": {"input":["text"]},
- "qwen/qwen3-next-80b-a3b-thinking": {"reasoning":true,"input":["text"]},
- "qwen/qwen3-vl-235b-a22b-instruct": {"input":["text","image"]},
- "qwen/qwen3-vl-235b-a22b-thinking": {"reasoning":true,"input":["text","image"]},
- "qwen/qwen3-vl-30b-a3b-instruct": {"input":["text","image"]},
- "qwen/qwen3-vl-30b-a3b-thinking": {"reasoning":true,"input":["text","image"]},
- "qwen/qwen3-vl-32b-instruct": {"input":["text","image"]},
- "qwen/qwen3-vl-8b-instruct": {"input":["image","text"]},
- "qwen/qwen3-vl-8b-thinking": {"reasoning":true,"input":["image","text"]},
- "qwen/qwen3.5-122b-a10b": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen/qwen3.5-27b": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen/qwen3.5-35b-a3b": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen/qwen3.5-397b-a17b": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen/qwen3.5-9b": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen/qwen3.5-flash-02-23": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen/qwen3.5-plus-02-15": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen/qwen3.5-plus-20260420": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen/qwen3.6-27b": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen/qwen3.6-35b-a3b": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen/qwen3.6-flash": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen/qwen3.6-max-preview": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen/qwen3.6-plus": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen/qwen3.7-flash": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen/qwen3.7-max": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen/qwen3.7-plus": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "qwen/qwen3.8-2.4t-a95b": {"reasoning":true,"efforts":["low","medium","xhigh"],"input":["text"]},
- "qwen/qwen3.8-27b": {"reasoning":true,"efforts":["low","medium","xhigh"],"toggle":true,"input":["text","image","video"]},
- "qwen/qwen3.8-flash": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen/qwen3.8-max-0902": {"reasoning":true,"efforts":["minimal","low","medium","high","xhigh"],"input":["text","image","video"]},
- "rekaai/reka-edge": {"input":["image","text","video"]},
- "rekaai/reka-flash-3": {"reasoning":true,"input":["text"]},
- "relace/relace-apply-3": {"input":["text"]},
- "relace/relace-search": {"input":["text"]},
- "sakana/fugu-max": {"reasoning":true,"efforts":["high","xhigh","max"],"input":["text","image","pdf"]},
- "sakana/fugu-ultra": {"reasoning":true,"efforts":["high","xhigh","max"],"input":["text","image"]},
- "sakana/fugu-ultra-v2": {"reasoning":true,"efforts":["high","xhigh","max"],"input":["text","image","pdf"]},
- "sakana/sakana-namazu": {"reasoning":true,"efforts":["none","high"],"input":["text","image","pdf"]},
- "sao10k/l3-lunaris-8b": {"input":["text"]},
- "sao10k/l3.1-euryale-70b": {"input":["text"]},
- "sao10k/l3.3-euryale-70b": {"input":["text"]},
- "stepfun/step-3.5-flash": {"reasoning":true,"input":["text"]},
- "stepfun/step-3.7-flash": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video"]},
- "tencent/hunyuan-a13b-instruct": {"reasoning":true,"toggle":true,"input":["text"]},
- "tencent/hy-mt2-1.8b": {"input":["text"]},
- "tencent/hy-mt2-30b-a3b": {"input":["text"]},
- "tencent/hy-mt2-7b": {"input":["text"]},
- "tencent/hy3": {"reasoning":true,"efforts":["none","low","high"],"input":["text"]},
- "tencent/hy3-preview": {"reasoning":true,"efforts":["none","low","high"],"input":["text"]},
- "tencent/hy4-preview": {"reasoning":true,"efforts":["none","low","high"],"input":["text"]},
- "thedrummer/cydonia-24b-v4.1": {"input":["text"]},
- "thedrummer/skyfall-36b-v2": {"input":["text"]},
- "thedrummer/unslopnemo-12b": {"input":["text"]},
- "thinkingmachines/inkling": {"reasoning":true,"efforts":["none","minimal","low","medium","high","max"],"input":["text","image","audio"]},
- "thinkingmachines/inkling-small": {"reasoning":true,"efforts":["none","minimal","low","medium","high","max"],"input":["text","image","audio"]},
- "thinkingmachines/inkling-small:free": {"reasoning":true,"efforts":["none","minimal","low","medium","high","max"],"input":["text","image","audio"]},
- "thinkingmachines/inkling:free": {"reasoning":true,"efforts":["none","minimal","low","medium","high","max"],"input":["text","image","audio"]},
- "undi95/remm-slerp-l2-13b": {"input":["text"]},
- "upstage/solar-pro-3": {"reasoning":true,"toggle":true,"input":["text"]},
- "upstage/solar-pro4": {"reasoning":true,"toggle":true,"input":["text"]},
- "writer/palmyra-x5": {"input":["text"]},
- "x-ai/grok-4.20": {"reasoning":true,"toggle":true,"input":["text","image","pdf"]},
- "x-ai/grok-4.20-multi-agent": {"reasoning":true,"efforts":["low","medium","high","xhigh"],"input":["text","image","pdf"]},
- "x-ai/grok-4.3": {"reasoning":true,"efforts":["none","low","medium","high"],"input":["text","image","pdf"]},
- "x-ai/grok-4.5": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","pdf"]},
- "x-ai/grok-4.6": {"reasoning":true,"efforts":["low","medium","high","xhigh"],"input":["text","image","pdf"]},
- "x-ai/grok-build-0.1": {"reasoning":true,"input":["text","image","pdf"]},
- "xiaomi/mimo-v2.5": {"reasoning":true,"toggle":true,"input":["text","image","audio","video"]},
- "xiaomi/mimo-v2.5-pro": {"reasoning":true,"toggle":true,"input":["text"]},
- "z-ai/glm-4.5": {"reasoning":true,"toggle":true,"input":["text"]},
- "z-ai/glm-4.5-air": {"reasoning":true,"toggle":true,"input":["text"]},
- "z-ai/glm-4.5v": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "z-ai/glm-4.6": {"reasoning":true,"toggle":true,"input":["text"]},
- "z-ai/glm-4.6v": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "z-ai/glm-4.7": {"reasoning":true,"toggle":true,"input":["text"]},
- "z-ai/glm-4.7-flash": {"reasoning":true,"toggle":true,"input":["text"]},
- "z-ai/glm-5": {"reasoning":true,"toggle":true,"input":["text"]},
- "z-ai/glm-5-turbo": {"reasoning":true,"toggle":true,"input":["text"]},
- "z-ai/glm-5.1": {"reasoning":true,"toggle":true,"input":["text"]},
- "z-ai/glm-5.2": {"reasoning":true,"efforts":["high","xhigh"],"toggle":true,"input":["text"]},
- "z-ai/glm-5.3": {"reasoning":true,"efforts":["low","high","max"],"input":["text"]},
- "z-ai/glm-5.3-flash": {"reasoning":true,"efforts":["low","high","max"],"input":["text","image","video"]},
- "z-ai/glm-5v-turbo": {"reasoning":true,"toggle":true,"input":["image","text","video"]},
+ "~anthropic/claude-fable-latest": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1000000},
+ "~anthropic/claude-haiku-latest": {"reasoning":true,"toggle":true,"input":["text","image","pdf"],"context":200000},
+ "~anthropic/claude-opus-latest": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"toggle":true,"input":["text","image","pdf"],"context":1000000},
+ "~anthropic/claude-sonnet-latest": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"toggle":true,"input":["text","image","pdf"],"context":1000000},
+ "~deepseek/deepseek-flash-latest": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image"],"context":1048576},
+ "~deepseek/deepseek-pro-latest": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text"],"context":1048576},
+ "~deepseek/deepseek-v4-flash-latest": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text"],"context":1310720},
+ "~google/gemini-flash-latest": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","pdf","audio"],"context":1048576},
+ "~google/gemini-pro-latest": {"reasoning":true,"efforts":["low","medium","high"],"input":["audio","pdf","image","text","video"],"context":1048576},
+ "~moonshotai/kimi-latest": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image","video"],"context":1048576},
+ "~openai/gpt-astra-latest": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["pdf","image","text"],"context":1050000},
+ "~openai/gpt-luna-latest": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["pdf","image","text"],"context":1050000},
+ "~openai/gpt-mini-latest": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["pdf","image","text"],"context":400000},
+ "~openai/gpt-sol-latest": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["pdf","image","text"],"context":1050000},
+ "~openai/gpt-terra-latest": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["pdf","image","text"],"context":1050000},
+ "~x-ai/grok-latest": {"reasoning":true,"efforts":["low","medium","high","xhigh"],"input":["text","image","pdf"],"context":500000},
+ "~z-ai/glm-flash-latest": {"reasoning":true,"efforts":["low","high","max"],"input":["text","image","video"],"context":1310720},
+ "~z-ai/glm-latest": {"reasoning":true,"efforts":["low","high","max"],"input":["text"],"context":1310720},
+ "aion-labs/aion-2.0": {"reasoning":true,"input":["text"],"context":131072},
+ "aion-labs/aion-3.0": {"reasoning":true,"input":["text"],"context":131072},
+ "aion-labs/aion-3.0-mini": {"reasoning":true,"input":["text"],"context":131072},
+ "aion-labs/aion-rp-llama-3.1-8b": {"input":["text"],"context":32768},
+ "amazon/nova-2-lite-v1": {"reasoning":true,"toggle":true,"input":["text","image","video","pdf"],"context":1000000},
+ "amazon/nova-lite-v1": {"input":["text","image"],"context":300000},
+ "amazon/nova-micro-v1": {"input":["text"],"context":128000},
+ "amazon/nova-premier-v1": {"input":["text","image"],"context":1000000},
+ "amazon/nova-pro-v1": {"input":["text","image"],"context":300000},
+ "anthracite-org/magnum-v4-72b": {"input":["text"],"context":32768},
+ "anthropic/claude-3-haiku": {"input":["text","image"],"context":200000},
+ "anthropic/claude-fable-5": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1000000},
+ "anthropic/claude-fable-5.1": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1000000},
+ "anthropic/claude-haiku-4.5": {"reasoning":true,"toggle":true,"input":["text","image","pdf"],"context":200000},
+ "anthropic/claude-opus-4": {"reasoning":true,"toggle":true,"input":["image","text","pdf"],"context":200000},
+ "anthropic/claude-opus-4.1": {"reasoning":true,"toggle":true,"input":["text","image","pdf"],"context":200000},
+ "anthropic/claude-opus-4.5": {"reasoning":true,"toggle":true,"input":["text","image","pdf"],"context":200000},
+ "anthropic/claude-opus-4.6": {"reasoning":true,"efforts":["low","medium","high","max"],"toggle":true,"input":["text","image","pdf"],"context":1000000},
+ "anthropic/claude-opus-4.7": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"toggle":true,"input":["text","image","pdf"],"context":1000000},
+ "anthropic/claude-opus-4.8": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"toggle":true,"input":["text","image","pdf"],"context":1000000},
+ "anthropic/claude-opus-5": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"toggle":true,"input":["text","image","pdf"],"context":1000000},
+ "anthropic/claude-sonnet-4": {"reasoning":true,"toggle":true,"input":["image","text","pdf"],"context":1000000},
+ "anthropic/claude-sonnet-4.5": {"reasoning":true,"toggle":true,"input":["text","image","pdf"],"context":1000000},
+ "anthropic/claude-sonnet-4.6": {"reasoning":true,"efforts":["low","medium","high","max"],"toggle":true,"input":["text","image","pdf"],"context":1000000},
+ "anthropic/claude-sonnet-5": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"toggle":true,"input":["text","image","pdf"],"context":1000000},
+ "arcee-ai/trinity-large-thinking": {"reasoning":true,"input":["text"],"context":262144},
+ "baidu/ernie-4.5-vl-424b-a47b": {"reasoning":true,"toggle":true,"input":["image","text"],"context":123000},
+ "bytedance-seed/seed-1.6": {"reasoning":true,"toggle":true,"input":["image","text","video"],"context":262144},
+ "bytedance-seed/seed-1.6-flash": {"reasoning":true,"toggle":true,"input":["image","text","video"],"context":262144},
+ "bytedance-seed/seed-2-1-turbo": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":262144},
+ "bytedance-seed/seed-2.0-code": {"reasoning":true,"efforts":["low","medium","high"],"toggle":true,"input":["text","image","video"],"context":262144},
+ "bytedance-seed/seed-2.0-lite": {"reasoning":true,"efforts":["minimal","low","medium","high"],"toggle":true,"input":["text","image","video"],"context":262144},
+ "bytedance-seed/seed-2.0-mini": {"reasoning":true,"efforts":["minimal","low","medium","high"],"toggle":true,"input":["text","image","video"],"context":262144},
+ "bytedance/ui-tars-1.5-7b": {"input":["image","text"],"context":128000},
+ "cognitivecomputations/dolphin-mistral-24b-venice-edition": {"input":["text"],"context":128000},
+ "cohere/command-a": {"input":["text"],"context":256000},
+ "cohere/command-r-08-2024": {"input":["text"],"context":128000},
+ "cohere/command-r-plus-08-2024": {"input":["text"],"context":128000},
+ "cohere/command-r7b-12-2024": {"input":["text"],"context":128000},
+ "cohere/north-mini-code:free": {"reasoning":true,"toggle":true,"input":["text"],"context":256000},
+ "deepseek/deepseek-chat": {"input":["text"],"context":163840},
+ "deepseek/deepseek-chat-v3-0324": {"input":["text"],"context":163840},
+ "deepseek/deepseek-chat-v3.1": {"reasoning":true,"toggle":true,"input":["text"],"context":163840},
+ "deepseek/deepseek-r1": {"reasoning":true,"input":["text"],"context":64000},
+ "deepseek/deepseek-r1-0528": {"reasoning":true,"input":["text"],"context":163840},
+ "deepseek/deepseek-r1-distill-llama-70b": {"reasoning":true,"toggle":true,"input":["text"],"context":8192},
+ "deepseek/deepseek-v3.1-terminus": {"reasoning":true,"toggle":true,"input":["text"],"context":163840},
+ "deepseek/deepseek-v3.2": {"reasoning":true,"toggle":true,"input":["text"],"context":163840},
+ "deepseek/deepseek-v3.2-exp": {"reasoning":true,"toggle":true,"input":["text"],"context":163840},
+ "deepseek/deepseek-v4-flash": {"reasoning":true,"efforts":["high","xhigh"],"toggle":true,"input":["text"],"context":1048576},
+ "deepseek/deepseek-v4-flash-0731": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text"],"context":1310720},
+ "deepseek/deepseek-v4-flash-vision-exp": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image"],"context":1048576},
+ "deepseek/deepseek-v4-pro": {"reasoning":true,"efforts":["high","xhigh"],"toggle":true,"input":["text"],"context":1048576},
+ "deepseek/deepseek-v4-pro-0813": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text"],"context":1048576},
+ "deepseek/deepseek-v4.1-flash": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image"],"context":1048576},
+ "dots-studio/dots-3-note-preview:free": {"reasoning":true,"toggle":true,"input":["text","image"],"context":512000},
+ "google/gemini-2.5-flash": {"reasoning":true,"toggle":true,"input":["text","image","audio","video","pdf"],"context":1048576},
+ "google/gemini-2.5-flash-image": {"input":["text","image"],"context":32768},
+ "google/gemini-2.5-flash-lite": {"reasoning":true,"toggle":true,"input":["text","image","audio","video","pdf"],"context":1048576},
+ "google/gemini-2.5-pro": {"reasoning":true,"input":["text","image","audio","video","pdf"],"context":1048576},
+ "google/gemini-2.5-pro-preview": {"reasoning":true,"input":["pdf","image","text","audio"],"context":1048576},
+ "google/gemini-3-flash-preview": {"reasoning":true,"efforts":["minimal","low","medium","high"],"toggle":true,"input":["text","image","video","audio","pdf"],"context":1048576},
+ "google/gemini-3-pro-image": {"reasoning":true,"input":["text","image"],"context":131072},
+ "google/gemini-3-pro-image-preview": {"reasoning":true,"input":["text","image"],"context":65536},
+ "google/gemini-3.1-flash-image": {"reasoning":true,"efforts":["minimal","high"],"toggle":true,"input":["image","text"],"context":131072},
+ "google/gemini-3.1-flash-image-preview": {"reasoning":true,"efforts":["minimal","high"],"toggle":true,"input":["image","text"],"context":65536},
+ "google/gemini-3.1-flash-lite": {"reasoning":true,"efforts":["minimal","low","medium","high"],"toggle":true,"input":["text","image","video","audio","pdf"],"context":1048576},
+ "google/gemini-3.1-flash-lite-image": {"reasoning":true,"efforts":["minimal","high"],"toggle":true,"input":["text","image"],"context":65536},
+ "google/gemini-3.1-flash-lite-preview": {"reasoning":true,"efforts":["minimal","low","medium","high"],"toggle":true,"input":["text","image","video","audio","pdf"],"context":1048576},
+ "google/gemini-3.1-pro-preview": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "google/gemini-3.1-pro-preview-customtools": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "google/gemini-3.5-flash": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "google/gemini-3.5-flash-lite": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "google/gemini-3.6-flash": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "google/gemini-3.7-flash": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "google/gemini-3.8-flash": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "google/gemma-2-27b-it": {"input":["text"],"context":8192},
+ "google/gemma-3-12b-it": {"input":["text","image"],"context":131072},
+ "google/gemma-3-27b-it": {"input":["text","image"],"context":131072},
+ "google/gemma-3-4b-it": {"input":["text","image"],"context":131072},
+ "google/gemma-4-26b-a4b-it": {"reasoning":true,"toggle":true,"input":["image","text","video"],"context":262144},
+ "google/gemma-4-26b-a4b-it:free": {"reasoning":true,"toggle":true,"input":["image","text","video"],"context":262144},
+ "google/gemma-4-31b-it": {"reasoning":true,"toggle":true,"input":["image","text","video"],"context":262144},
+ "google/gemma-4-31b-it:free": {"reasoning":true,"toggle":true,"input":["image","text","video"],"context":262144},
+ "google/lyria-3-clip-preview": {"input":["text","image"],"context":1048576},
+ "google/lyria-3-pro-preview": {"input":["text","image"],"context":1048576},
+ "gryphe/mythomax-l2-13b": {"input":["text"],"context":8192},
+ "ibm-granite/granite-4.0-h-micro": {"input":["text"],"context":131000},
+ "ibm-granite/granite-4.2-8b": {"reasoning":true,"efforts":["none","low","high"],"input":["text"],"context":131072},
+ "inception/mercury-2": {"reasoning":true,"efforts":["none","low","medium","high"],"input":["text"],"context":128000},
+ "inception/mercury-2.5": {"reasoning":true,"efforts":["none","low","medium","high"],"input":["text"],"context":260000},
+ "inclusionai/ling-3.0-flash": {"reasoning":true,"toggle":true,"input":["text"],"context":262144},
+ "inclusionai/ling-3.0-flash-fin": {"reasoning":true,"toggle":true,"input":["text"],"context":262144},
+ "inclusionai/ling-3.0-flash-fin:free": {"reasoning":true,"toggle":true,"input":["text"],"context":262144},
+ "inclusionai/ling-3.0-flash-sante:free": {"reasoning":true,"toggle":true,"input":["text"],"context":262144},
+ "inclusionai/ling-3.0-flash-vl": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":131072},
+ "inclusionai/ling-3.0-flash-vl:free": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":262144},
+ "inference-net/schematron-v2-small": {"input":["text"],"context":128000},
+ "inference-net/schematron-v2-turbo": {"input":["text"],"context":128000},
+ "kwaipilot/kat-coder-pro-v2": {"input":["text"],"context":262144},
+ "kwaipilot/kat-coder-pro-v2.5": {"input":["text"],"context":262144},
+ "liquid/lfm-2.5-2.6b:free": {"reasoning":true,"input":["text"],"context":65536},
+ "mancer/weaver": {"input":["text"],"context":8000},
+ "meituan/longcat-2.0": {"reasoning":true,"toggle":true,"input":["text"],"context":1048756},
+ "meta-llama/llama-3.1-70b-instruct": {"input":["text"],"context":131072},
+ "meta-llama/llama-3.1-8b-instruct": {"input":["text"],"context":131072},
+ "meta-llama/llama-3.2-1b-instruct": {"input":["text"],"context":60000},
+ "meta-llama/llama-3.2-3b-instruct": {"input":["text"],"context":131072},
+ "meta-llama/llama-3.3-70b-instruct": {"input":["text"],"context":131072},
+ "meta-llama/llama-4-maverick": {"input":["text","image"],"context":1048576},
+ "meta-llama/llama-4-scout": {"input":["text","image"],"context":1310720},
+ "meta-llama/llama-guard-4-12b": {"input":["image","text"],"context":163840},
+ "meta/muse-glimmer-30b": {"reasoning":true,"efforts":["low","medium","high","xhigh"],"input":["text","image"],"context":131072},
+ "meta/muse-spark-1.1": {"reasoning":true,"efforts":["minimal","low","medium","high","xhigh"],"input":["text","image","video","pdf","audio"],"context":1048576},
+ "meta/muse-spark-1.2": {"reasoning":true,"efforts":["minimal","low","medium","high","xhigh"],"input":["text","image","video","pdf","audio"],"context":1048576},
+ "meta/muse-spark-1.2-contributor": {"reasoning":true,"efforts":["minimal","low","medium","high","xhigh"],"input":["text","image","video","pdf","audio"],"context":1048576},
+ "meta/muse-spark-1.3": {"reasoning":true,"efforts":["minimal","low","medium","high","xhigh","max"],"input":["text","image","video","pdf","audio"],"context":1048576},
+ "meta/muse-spark-1.3-contributor": {"reasoning":true,"efforts":["minimal","low","medium","high","xhigh","max"],"input":["text","image","video","pdf","audio"],"context":1048576},
+ "microsoft/phi-4": {"input":["text"],"context":16384},
+ "microsoft/wizardlm-2-8x22b": {"input":["text"],"context":65535},
+ "minimax/minimax-01": {"input":["text","image"],"context":1000192},
+ "minimax/minimax-m1": {"reasoning":true,"toggle":true,"input":["text"],"context":1000000},
+ "minimax/minimax-m2": {"reasoning":true,"input":["text"],"context":204800},
+ "minimax/minimax-m2-her": {"input":["text"],"context":65536},
+ "minimax/minimax-m2.1": {"reasoning":true,"input":["text"],"context":204800},
+ "minimax/minimax-m2.5": {"reasoning":true,"input":["text"],"context":204800},
+ "minimax/minimax-m2.7": {"reasoning":true,"input":["text"],"context":204800},
+ "minimax/minimax-m3": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":1048576},
+ "mistralai/codestral-2508": {"input":["text","pdf"],"context":256000},
+ "mistralai/devstral-2512": {"input":["text","pdf"],"context":262144},
+ "mistralai/ministral-14b-2512": {"input":["text","image"],"context":262144},
+ "mistralai/ministral-3b-2512": {"input":["text","image"],"context":131072},
+ "mistralai/ministral-8b-2512": {"input":["text","image"],"context":262144},
+ "mistralai/mistral-large": {"input":["text","pdf"],"context":128000},
+ "mistralai/mistral-large-2407": {"input":["text","pdf"],"context":131072},
+ "mistralai/mistral-large-2512": {"input":["text","image","pdf"],"context":262144},
+ "mistralai/mistral-medium-3": {"input":["text","image","pdf"],"context":131072},
+ "mistralai/mistral-medium-3-5": {"reasoning":true,"efforts":["none","high"],"input":["text","image","pdf"],"context":262144},
+ "mistralai/mistral-medium-3.1": {"input":["text","image","pdf"],"context":131072},
+ "mistralai/mistral-nemo": {"input":["text"],"context":131072},
+ "mistralai/mistral-saba": {"input":["text","pdf"],"context":32768},
+ "mistralai/mistral-small-24b-instruct-2501": {"input":["text"],"context":32768},
+ "mistralai/mistral-small-2603": {"reasoning":true,"efforts":["none","high"],"input":["text","image"],"context":262144},
+ "mistralai/mistral-small-3.1-24b-instruct": {"input":["text","image"],"context":128000},
+ "mistralai/mistral-small-3.2-24b-instruct": {"input":["image","text"],"context":256000},
+ "mistralai/mixtral-8x22b-instruct": {"input":["text","pdf"],"context":65536},
+ "mistralai/voxtral-small-24b-2507": {"input":["text","audio","pdf"],"context":32768},
+ "moonshotai/kimi-k2": {"input":["text"],"context":131072},
+ "moonshotai/kimi-k2-0905": {"input":["text"],"context":262144},
+ "moonshotai/kimi-k2-thinking": {"reasoning":true,"input":["text"],"context":262144},
+ "moonshotai/kimi-k2.5": {"reasoning":true,"toggle":true,"input":["text","image"],"context":262144},
+ "moonshotai/kimi-k2.6": {"reasoning":true,"toggle":true,"input":["text","image"],"context":262144},
+ "moonshotai/kimi-k2.7-code": {"reasoning":true,"input":["text","image"],"context":262144},
+ "moonshotai/kimi-k3": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image","video"],"context":1048576},
+ "morph/morph-v3-fast": {"input":["text"],"context":81920},
+ "morph/morph-v3-large": {"input":["text"],"context":262144},
+ "nex-agi/nex-n2.5-mini:free": {"reasoning":true,"efforts":["none","medium","high"],"input":["text","image"],"context":262144},
+ "nex-agi/nex-n2.5-pro:free": {"reasoning":true,"efforts":["none","medium","high"],"input":["text","image"],"context":262144},
+ "nousresearch/hermes-3-llama-3.1-405b": {"input":["text"],"context":131072},
+ "nousresearch/hermes-3-llama-3.1-70b": {"input":["text"],"context":131072},
+ "nousresearch/hermes-4-405b": {"reasoning":true,"toggle":true,"input":["text"],"context":131072},
+ "nvidia/nemotron-3-nano-30b-a3b": {"reasoning":true,"toggle":true,"input":["text"],"context":262144},
+ "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free": {"reasoning":true,"toggle":true,"input":["text","image","video","audio"],"context":256000},
+ "nvidia/nemotron-3-super-120b-a12b": {"reasoning":true,"efforts":["low","medium"],"toggle":true,"input":["text"],"context":262144},
+ "nvidia/nemotron-3-super-120b-a12b:free": {"reasoning":true,"efforts":["low","medium"],"toggle":true,"input":["text"],"context":262144},
+ "nvidia/nemotron-3-ultra-550b-a55b": {"reasoning":true,"efforts":["medium","high"],"toggle":true,"input":["text"],"context":262144},
+ "nvidia/nemotron-3-ultra-550b-a55b:free": {"reasoning":true,"efforts":["medium","high"],"toggle":true,"input":["text"],"context":1000000},
+ "nvidia/nemotron-3.5-content-safety": {"reasoning":true,"toggle":true,"input":["text","image"],"context":131072},
+ "nvidia/nemotron-3.5-content-safety:free": {"reasoning":true,"toggle":true,"input":["text","image"],"context":128000},
+ "nvidia/nemotron-3.5-lightning": {"reasoning":true,"toggle":true,"input":["text"],"context":262144},
+ "nvidia/nemotron-3.5-lightning:free": {"reasoning":true,"toggle":true,"input":["text"],"context":1000000},
+ "openai/gpt-3.5-turbo": {"input":["text"],"context":16385},
+ "openai/gpt-3.5-turbo-0613": {"input":["text"],"context":4095},
+ "openai/gpt-3.5-turbo-16k": {"input":["text"],"context":16385},
+ "openai/gpt-3.5-turbo-instruct": {"input":["text"],"context":4095},
+ "openai/gpt-4": {"input":["text"],"context":8191},
+ "openai/gpt-4-turbo": {"input":["text","image"],"context":128000},
+ "openai/gpt-4.1": {"input":["text","image","pdf"],"context":1047576},
+ "openai/gpt-4.1-mini": {"input":["text","image","pdf"],"context":1047576},
+ "openai/gpt-4.1-nano": {"input":["image","text","pdf"],"context":1047576},
+ "openai/gpt-4o": {"input":["text","image","pdf"],"context":128000},
+ "openai/gpt-4o-2024-05-13": {"input":["text","image","pdf"],"context":128000},
+ "openai/gpt-4o-2024-08-06": {"input":["text","image","pdf"],"context":128000},
+ "openai/gpt-4o-2024-11-20": {"input":["text","image","pdf"],"context":128000},
+ "openai/gpt-4o-mini": {"input":["text","image","pdf"],"context":128000},
+ "openai/gpt-4o-mini-2024-07-18": {"input":["text","image","pdf"],"context":128000},
+ "openai/gpt-5": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","pdf"],"context":400000},
+ "openai/gpt-5-image": {"reasoning":true,"input":["image","text","pdf"],"context":400000},
+ "openai/gpt-5-image-mini": {"reasoning":true,"input":["pdf","image","text"],"context":400000},
+ "openai/gpt-5-mini": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","pdf"],"context":400000},
+ "openai/gpt-5-nano": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","pdf"],"context":400000},
+ "openai/gpt-5-pro": {"reasoning":true,"efforts":["high"],"input":["image","text","pdf"],"context":400000},
+ "openai/gpt-5.1": {"reasoning":true,"efforts":["none","low","medium","high"],"input":["image","text","pdf"],"context":400000},
+ "openai/gpt-5.1-codex": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image"],"context":400000},
+ "openai/gpt-5.1-codex-max": {"reasoning":true,"efforts":["low","medium","high","xhigh"],"input":["text","image"],"context":400000},
+ "openai/gpt-5.1-codex-mini": {"reasoning":true,"efforts":["low","medium","high"],"toggle":true,"input":["text","image"],"context":400000},
+ "openai/gpt-5.2": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["pdf","image","text"],"context":400000},
+ "openai/gpt-5.2-chat": {"input":["pdf","image","text"],"context":128000},
+ "openai/gpt-5.2-codex": {"reasoning":true,"efforts":["low","medium","high","xhigh"],"input":["text","image"],"context":400000},
+ "openai/gpt-5.2-pro": {"reasoning":true,"efforts":["medium","high","xhigh"],"input":["image","text","pdf"],"context":400000},
+ "openai/gpt-5.3-codex": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image","pdf"],"context":400000},
+ "openai/gpt-5.4": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image","pdf"],"context":1050000},
+ "openai/gpt-5.4-image-2": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["image","text","pdf"],"context":272000},
+ "openai/gpt-5.4-mini": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["pdf","image","text"],"context":400000},
+ "openai/gpt-5.4-nano": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["pdf","image","text"],"context":400000},
+ "openai/gpt-5.4-pro": {"reasoning":true,"efforts":["medium","high","xhigh"],"input":["text","image","pdf"],"context":1050000},
+ "openai/gpt-5.5": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"input":["text","image","pdf"],"context":1050000},
+ "openai/gpt-5.5-pro": {"reasoning":true,"efforts":["medium","high","xhigh"],"input":["text","image","pdf"],"context":1050000},
+ "openai/gpt-5.6-luna": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1050000},
+ "openai/gpt-5.6-luna-pro": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1050000},
+ "openai/gpt-5.6-sol": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1050000},
+ "openai/gpt-5.6-sol-pro": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1050000},
+ "openai/gpt-5.6-terra": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1050000},
+ "openai/gpt-5.6-terra-pro": {"reasoning":true,"efforts":["none","low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1050000},
+ "openai/gpt-6-astra": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["text","image","pdf"],"context":1050000},
+ "openai/gpt-6-astra-pro": {"reasoning":true,"efforts":["low","medium","high","xhigh","max"],"input":["pdf","image","text"],"context":1050000},
+ "openai/gpt-audio": {"input":["text","audio"],"context":128000},
+ "openai/gpt-audio-mini": {"input":["text","audio"],"context":128000},
+ "openai/gpt-chat-latest": {"input":["text","image","pdf"],"context":400000},
+ "openai/gpt-oss-120b": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"],"context":131072},
+ "openai/gpt-oss-20b": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"],"context":131072},
+ "openai/gpt-oss-safeguard-20b": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"],"context":131072},
+ "openai/o1": {"reasoning":true,"toggle":true,"input":["text","image","pdf"],"context":200000},
+ "openai/o1-pro": {"reasoning":true,"toggle":true,"input":["text","image","pdf"],"context":200000},
+ "openai/o3": {"reasoning":true,"toggle":true,"input":["text","image","pdf"],"context":200000},
+ "openai/o3-mini": {"reasoning":true,"toggle":true,"input":["text","pdf"],"context":200000},
+ "openai/o3-mini-high": {"reasoning":true,"efforts":["high"],"input":["text","pdf"],"context":200000},
+ "openai/o3-pro": {"reasoning":true,"toggle":true,"input":["text","pdf","image"],"context":200000},
+ "openai/o4-mini": {"reasoning":true,"toggle":true,"input":["image","text","pdf"],"context":200000},
+ "openai/o4-mini-high": {"reasoning":true,"efforts":["high"],"input":["image","text","pdf"],"context":200000},
+ "openrouter/auto": {"reasoning":true,"input":["text","image","audio","pdf","video"],"context":2000000},
+ "openrouter/bodybuilder": {"input":["text"],"context":128000},
+ "openrouter/free": {"reasoning":true,"input":["text","image"],"context":200000},
+ "openrouter/fusion": {"input":["text"],"context":1000000},
+ "openrouter/pareto-code": {"input":["text"],"context":2000000},
+ "perceptron/perceptron-mk1": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":32768},
+ "perplexity/sonar": {"input":["text","image"],"context":127072},
+ "perplexity/sonar-deep-research": {"reasoning":true,"toggle":true,"input":["text"],"context":128000},
+ "perplexity/sonar-pro": {"input":["text","image"],"context":200000},
+ "perplexity/sonar-pro-search": {"reasoning":true,"input":["text","image"],"context":200000},
+ "perplexity/sonar-reasoning-pro": {"reasoning":true,"toggle":true,"input":["text","image"],"context":128000},
+ "poolside/laguna-s-2.1": {"reasoning":true,"toggle":true,"input":["text"],"context":1048576},
+ "poolside/laguna-s-2.1:free": {"reasoning":true,"toggle":true,"input":["text"],"context":262144},
+ "poolside/laguna-xs-2.1": {"reasoning":true,"toggle":true,"input":["text"],"context":262144},
+ "poolside/laguna-xs-2.1:free": {"reasoning":true,"toggle":true,"input":["text"],"context":262144},
+ "qwen/qwen-2.5-72b-instruct": {"input":["text"],"context":32768},
+ "qwen/qwen-2.5-7b-instruct": {"input":["text"],"context":32768},
+ "qwen/qwen-2.5-coder-32b-instruct": {"input":["text"],"context":32768},
+ "qwen/qwen-plus": {"input":["text"],"context":1000000},
+ "qwen/qwen-plus-2025-07-28": {"input":["text"],"context":1000000},
+ "qwen/qwen2.5-vl-72b-instruct": {"input":["text","image"],"context":128000},
+ "qwen/qwen3-14b": {"reasoning":true,"toggle":true,"input":["text"],"context":131072},
+ "qwen/qwen3-235b-a22b": {"reasoning":true,"toggle":true,"input":["text"],"context":131072},
+ "qwen/qwen3-235b-a22b-2507": {"input":["text"],"context":262144},
+ "qwen/qwen3-235b-a22b-thinking-2507": {"reasoning":true,"input":["text"],"context":131072},
+ "qwen/qwen3-30b-a3b": {"reasoning":true,"toggle":true,"input":["text"],"context":131072},
+ "qwen/qwen3-30b-a3b-instruct-2507": {"input":["text"],"context":262144},
+ "qwen/qwen3-30b-a3b-thinking-2507": {"reasoning":true,"input":["text"],"context":81920},
+ "qwen/qwen3-32b": {"reasoning":true,"toggle":true,"input":["text"],"context":131072},
+ "qwen/qwen3-8b": {"reasoning":true,"toggle":true,"input":["text"],"context":131072},
+ "qwen/qwen3-coder": {"input":["text"],"context":262144},
+ "qwen/qwen3-coder-30b-a3b-instruct": {"input":["text"],"context":262144},
+ "qwen/qwen3-coder-flash": {"input":["text"],"context":1000000},
+ "qwen/qwen3-coder-next": {"input":["text"],"context":262144},
+ "qwen/qwen3-coder-plus": {"input":["text"],"context":1000000},
+ "qwen/qwen3-max": {"input":["text"],"context":262144},
+ "qwen/qwen3-max-thinking": {"reasoning":true,"toggle":true,"input":["text"],"context":262144},
+ "qwen/qwen3-next-80b-a3b-instruct": {"input":["text"],"context":262144},
+ "qwen/qwen3-next-80b-a3b-thinking": {"reasoning":true,"input":["text"],"context":262144},
+ "qwen/qwen3-vl-235b-a22b-instruct": {"input":["text","image"],"context":262144},
+ "qwen/qwen3-vl-235b-a22b-thinking": {"reasoning":true,"input":["text","image"],"context":131072},
+ "qwen/qwen3-vl-30b-a3b-instruct": {"input":["text","image"],"context":262144},
+ "qwen/qwen3-vl-30b-a3b-thinking": {"reasoning":true,"input":["text","image"],"context":262144},
+ "qwen/qwen3-vl-32b-instruct": {"input":["text","image"],"context":131072},
+ "qwen/qwen3-vl-8b-instruct": {"input":["image","text"],"context":262144},
+ "qwen/qwen3-vl-8b-thinking": {"reasoning":true,"input":["image","text"],"context":131072},
+ "qwen/qwen3.5-122b-a10b": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":262144},
+ "qwen/qwen3.5-27b": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":262144},
+ "qwen/qwen3.5-35b-a3b": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":262144},
+ "qwen/qwen3.5-397b-a17b": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":262144},
+ "qwen/qwen3.5-9b": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":262144},
+ "qwen/qwen3.5-flash-02-23": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":1000000},
+ "qwen/qwen3.5-plus-02-15": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":1000000},
+ "qwen/qwen3.5-plus-20260420": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":1000000},
+ "qwen/qwen3.6-27b": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":262144},
+ "qwen/qwen3.6-35b-a3b": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":262144},
+ "qwen/qwen3.6-flash": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":1000000},
+ "qwen/qwen3.6-max-preview": {"reasoning":true,"toggle":true,"input":["text"],"context":262144},
+ "qwen/qwen3.6-plus": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":1000000},
+ "qwen/qwen3.7-flash": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":1000000},
+ "qwen/qwen3.7-max": {"reasoning":true,"toggle":true,"input":["text"],"context":1000000},
+ "qwen/qwen3.7-plus": {"reasoning":true,"toggle":true,"input":["text","image"],"context":1000000},
+ "qwen/qwen3.8-2.4t-a95b": {"reasoning":true,"efforts":["low","medium","xhigh"],"input":["text"],"context":1048576},
+ "qwen/qwen3.8-27b": {"reasoning":true,"efforts":["low","medium","xhigh"],"toggle":true,"input":["text","image","video"],"context":1000000},
+ "qwen/qwen3.8-flash": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":1000000},
+ "qwen/qwen3.8-max-0902": {"reasoning":true,"efforts":["minimal","low","medium","high","xhigh"],"input":["text","image","video"],"context":1000000},
+ "rekaai/reka-edge": {"input":["image","text","video"],"context":16384},
+ "rekaai/reka-flash-3": {"reasoning":true,"input":["text"],"context":65536},
+ "relace/relace-apply-3": {"input":["text"],"context":256000},
+ "relace/relace-search": {"input":["text"],"context":256000},
+ "sakana/fugu-max": {"reasoning":true,"efforts":["high","xhigh","max"],"input":["text","image","pdf"],"context":1000000},
+ "sakana/fugu-ultra": {"reasoning":true,"efforts":["high","xhigh","max"],"input":["text","image"],"context":1000000},
+ "sakana/fugu-ultra-v2": {"reasoning":true,"efforts":["high","xhigh","max"],"input":["text","image","pdf"],"context":1000000},
+ "sakana/sakana-namazu": {"reasoning":true,"efforts":["none","high"],"input":["text","image","pdf"],"context":262144},
+ "sao10k/l3-lunaris-8b": {"input":["text"],"context":8192},
+ "sao10k/l3.1-euryale-70b": {"input":["text"],"context":131072},
+ "sao10k/l3.3-euryale-70b": {"input":["text"],"context":131072},
+ "stealth/union-alpha": {"input":["text","image"],"context":262144},
+ "stepfun/step-3.5-flash": {"reasoning":true,"input":["text"],"context":262144},
+ "stepfun/step-3.7-flash": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video"],"context":262144},
+ "tencent/hunyuan-a13b-instruct": {"reasoning":true,"toggle":true,"input":["text"],"context":131072},
+ "tencent/hy-mt2-1.8b": {"input":["text"],"context":8192},
+ "tencent/hy-mt2-30b-a3b": {"input":["text"],"context":8192},
+ "tencent/hy-mt2-7b": {"input":["text"],"context":8192},
+ "tencent/hy3": {"reasoning":true,"efforts":["none","low","high"],"input":["text"],"context":262144},
+ "tencent/hy3-preview": {"reasoning":true,"efforts":["none","low","high"],"input":["text"],"context":262144},
+ "tencent/hy4-preview": {"reasoning":true,"efforts":["none","low","high"],"input":["text"],"context":1048576},
+ "thedrummer/cydonia-24b-v4.1": {"input":["text"],"context":131072},
+ "thedrummer/skyfall-36b-v2": {"input":["text"],"context":32768},
+ "thedrummer/unslopnemo-12b": {"input":["text"],"context":1024000},
+ "thinkingmachines/inkling": {"reasoning":true,"efforts":["none","minimal","low","medium","high","max"],"input":["text","image","audio"],"context":1048576},
+ "thinkingmachines/inkling-small": {"reasoning":true,"efforts":["none","minimal","low","medium","high","max"],"input":["text","image","audio"],"context":1048576},
+ "thinkingmachines/inkling-small:free": {"reasoning":true,"efforts":["none","minimal","low","medium","high","max"],"input":["text","image","audio"],"context":1048576},
+ "thinkingmachines/inkling:free": {"reasoning":true,"efforts":["none","minimal","low","medium","high","max"],"input":["text","image","audio"],"context":1048576},
+ "undi95/remm-slerp-l2-13b": {"input":["text"],"context":6144},
+ "upstage/solar-pro-3": {"reasoning":true,"toggle":true,"input":["text"],"context":131072},
+ "upstage/solar-pro4": {"reasoning":true,"toggle":true,"input":["text"],"context":524288},
+ "writer/palmyra-x5": {"input":["text"],"context":1040000},
+ "x-ai/grok-4.20": {"reasoning":true,"toggle":true,"input":["text","image","pdf"],"context":2000000},
+ "x-ai/grok-4.20-multi-agent": {"reasoning":true,"efforts":["low","medium","high","xhigh"],"input":["text","image","pdf"],"context":2000000},
+ "x-ai/grok-4.3": {"reasoning":true,"efforts":["none","low","medium","high"],"input":["text","image","pdf"],"context":1000000},
+ "x-ai/grok-4.5": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","pdf"],"context":500000},
+ "x-ai/grok-4.6": {"reasoning":true,"efforts":["low","medium","high","xhigh"],"input":["text","image","pdf"],"context":500000},
+ "x-ai/grok-build-0.1": {"reasoning":true,"input":["text","image","pdf"],"context":256000},
+ "xiaomi/mimo-v2.5": {"reasoning":true,"toggle":true,"input":["text","image","audio","video"],"context":1050000},
+ "xiaomi/mimo-v2.5-pro": {"reasoning":true,"toggle":true,"input":["text"],"context":1050000},
+ "z-ai/glm-4.5": {"reasoning":true,"toggle":true,"input":["text"],"context":131072},
+ "z-ai/glm-4.5-air": {"reasoning":true,"toggle":true,"input":["text"],"context":131072},
+ "z-ai/glm-4.5v": {"reasoning":true,"toggle":true,"input":["text","image"],"context":65536},
+ "z-ai/glm-4.6": {"reasoning":true,"toggle":true,"input":["text"],"context":204800},
+ "z-ai/glm-4.6v": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":131072},
+ "z-ai/glm-4.7": {"reasoning":true,"toggle":true,"input":["text"],"context":204800},
+ "z-ai/glm-4.7-flash": {"reasoning":true,"toggle":true,"input":["text"],"context":200000},
+ "z-ai/glm-5": {"reasoning":true,"toggle":true,"input":["text"],"context":204800},
+ "z-ai/glm-5-turbo": {"reasoning":true,"toggle":true,"input":["text"],"context":202752},
+ "z-ai/glm-5.1": {"reasoning":true,"toggle":true,"input":["text"],"context":204800},
+ "z-ai/glm-5.2": {"reasoning":true,"efforts":["high","xhigh"],"toggle":true,"input":["text"],"context":1048576},
+ "z-ai/glm-5.2:free": {"reasoning":true,"efforts":["high","xhigh"],"toggle":true,"input":["text"],"context":32768},
+ "z-ai/glm-5.3": {"reasoning":true,"efforts":["low","high","max"],"input":["text"],"context":1310720},
+ "z-ai/glm-5.3-flash": {"reasoning":true,"efforts":["low","high","max"],"input":["text","image","video"],"context":1310720},
+ "z-ai/glm-5v-turbo": {"reasoning":true,"toggle":true,"input":["image","text","video"],"context":202752},
},
"gemini": {
- "deep-research-max-preview-04-2026": {"reasoning":true,"input":["text","image","video","audio","pdf"]},
- "deep-research-preview-04-2026": {"reasoning":true,"input":["text","image","video","audio","pdf"]},
- "gemini-2.5-computer-use-preview-10-2025": {"reasoning":true,"input":["text","image"]},
- "gemini-2.5-flash": {"reasoning":true,"toggle":true,"input":["text","image","audio","video","pdf"]},
- "gemini-2.5-flash-image": {"reasoning":true,"input":["text","image"]},
- "gemini-2.5-flash-lite": {"reasoning":true,"toggle":true,"input":["text","image","audio","video","pdf"]},
- "gemini-2.5-flash-preview-tts": {"input":["text"]},
- "gemini-2.5-pro": {"reasoning":true,"input":["text","image","audio","video","pdf"]},
- "gemini-2.5-pro-preview-tts": {"input":["text"]},
- "gemini-3-flash-preview": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "gemini-3-pro-image": {"reasoning":true,"efforts":["low","high"],"input":["text","image"]},
- "gemini-3-pro-image-preview": {"reasoning":true,"input":["text","image"]},
- "gemini-3.1-flash-image": {"reasoning":true,"efforts":["minimal","high"],"input":["text","image","video","pdf"]},
- "gemini-3.1-flash-image-preview": {"reasoning":true,"efforts":["minimal","high"],"input":["text","image","pdf"]},
- "gemini-3.1-flash-lite": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "gemini-3.1-flash-lite-image": {"reasoning":true,"efforts":["minimal","high"],"input":["text","image"]},
- "gemini-3.1-flash-lite-preview": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "gemini-3.1-flash-live-preview": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio"]},
- "gemini-3.1-flash-tts-preview": {"reasoning":true,"input":["text"]},
- "gemini-3.1-pro-preview": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "gemini-3.1-pro-preview-customtools": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "gemini-3.5-flash": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "gemini-3.5-flash-lite": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "gemini-3.5-live-translate-preview": {"input":["audio"]},
- "gemini-3.6-flash": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "gemini-3.7-flash": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "gemini-3.8-flash": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "gemini-embedding-001": {"input":["text"]},
- "gemini-embedding-2": {"input":["text","image","audio","video","pdf"]},
- "gemini-flash-latest": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "gemini-flash-lite-latest": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"]},
- "gemini-omni-flash-preview": {"reasoning":true,"input":["text","image","video"]},
- "gemma-4-26b-a4b-it": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "gemma-4-31b-it": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "lyria-3-clip-preview": {"input":["text","image"]},
- "lyria-3-pro-preview": {"input":["text","image"]},
- "veo-3.1-fast-generate-preview": {"input":["text","image","video"]},
- "veo-3.1-generate-preview": {"input":["text","image"]},
- "veo-3.1-lite-generate-preview": {"input":["text","image"]},
+ "deep-research-max-preview-04-2026": {"reasoning":true,"input":["text","image","video","audio","pdf"],"context":131072},
+ "deep-research-preview-04-2026": {"reasoning":true,"input":["text","image","video","audio","pdf"],"context":131072},
+ "gemini-2.5-computer-use-preview-10-2025": {"reasoning":true,"input":["text","image"],"context":131072},
+ "gemini-2.5-flash": {"reasoning":true,"toggle":true,"input":["text","image","audio","video","pdf"],"context":1048576},
+ "gemini-2.5-flash-image": {"reasoning":true,"input":["text","image"],"context":32768},
+ "gemini-2.5-flash-lite": {"reasoning":true,"toggle":true,"input":["text","image","audio","video","pdf"],"context":1048576},
+ "gemini-2.5-flash-preview-tts": {"input":["text"],"context":8192},
+ "gemini-2.5-pro": {"reasoning":true,"input":["text","image","audio","video","pdf"],"context":1048576},
+ "gemini-2.5-pro-preview-tts": {"input":["text"],"context":8192},
+ "gemini-3-flash-preview": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "gemini-3-pro-image": {"reasoning":true,"efforts":["low","high"],"input":["text","image"],"context":131072},
+ "gemini-3-pro-image-preview": {"reasoning":true,"input":["text","image"],"context":131072},
+ "gemini-3.1-flash-image": {"reasoning":true,"efforts":["minimal","high"],"input":["text","image","video","pdf"],"context":65536},
+ "gemini-3.1-flash-image-preview": {"reasoning":true,"efforts":["minimal","high"],"input":["text","image","pdf"],"context":65536},
+ "gemini-3.1-flash-lite": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "gemini-3.1-flash-lite-image": {"reasoning":true,"efforts":["minimal","high"],"input":["text","image"],"context":65536},
+ "gemini-3.1-flash-lite-preview": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "gemini-3.1-flash-live-preview": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio"],"context":131072},
+ "gemini-3.1-flash-tts-preview": {"reasoning":true,"input":["text"],"context":8192},
+ "gemini-3.1-pro-preview": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "gemini-3.1-pro-preview-customtools": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "gemini-3.5-flash": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "gemini-3.5-flash-lite": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "gemini-3.5-live-translate-preview": {"input":["audio"],"context":16384},
+ "gemini-3.6-flash": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "gemini-3.7-flash": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "gemini-3.8-flash": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "gemini-embedding-001": {"input":["text"],"context":2048},
+ "gemini-embedding-2": {"input":["text","image","audio","video","pdf"],"context":8192},
+ "gemini-flash-latest": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "gemini-flash-lite-latest": {"reasoning":true,"efforts":["minimal","low","medium","high"],"input":["text","image","video","audio","pdf"],"context":1048576},
+ "gemini-omni-flash-preview": {"reasoning":true,"input":["text","image","video"],"context":131072},
+ "gemma-4-26b-a4b-it": {"reasoning":true,"toggle":true,"input":["text","image"],"context":262144},
+ "gemma-4-31b-it": {"reasoning":true,"toggle":true,"input":["text","image"],"context":262144},
+ "lyria-3-clip-preview": {"input":["text","image"],"context":1048576},
+ "lyria-3-pro-preview": {"input":["text","image"],"context":1048576},
+ "veo-3.1-fast-generate-preview": {"input":["text","image","video"],"context":480},
+ "veo-3.1-generate-preview": {"input":["text","image"],"context":480},
+ "veo-3.1-lite-generate-preview": {"input":["text","image"],"context":480},
},
"mistral": {
- "codestral-latest": {"input":["text"]},
- "devstral-2512": {"input":["text"]},
- "devstral-latest": {"input":["text"]},
- "devstral-medium-2507": {"input":["text"]},
- "devstral-medium-latest": {"input":["text"]},
- "devstral-small-2505": {"input":["text"]},
- "devstral-small-2507": {"input":["text"]},
- "labs-devstral-small-2512": {"input":["text","image"]},
- "magistral-medium-latest": {"reasoning":true,"input":["text"]},
- "magistral-small": {"reasoning":true,"input":["text"]},
- "ministral-3b-latest": {"input":["text"]},
- "ministral-8b-latest": {"input":["text"]},
- "mistral-embed": {"input":["text"]},
- "mistral-large-2411": {"input":["text"]},
- "mistral-large-2512": {"input":["text","image"]},
- "mistral-large-latest": {"input":["text","image"]},
- "mistral-medium-2505": {"input":["text","image"]},
- "mistral-medium-2508": {"input":["text","image"]},
- "mistral-medium-2604": {"reasoning":true,"efforts":["none","high"],"input":["text","image"]},
- "mistral-medium-latest": {"reasoning":true,"efforts":["none","high"],"input":["text","image"]},
- "mistral-nemo": {"input":["text"]},
- "mistral-small-2506": {"input":["text","image"]},
- "mistral-small-2603": {"reasoning":true,"efforts":["none","high"],"input":["text","image"]},
- "mistral-small-latest": {"reasoning":true,"efforts":["none","high"],"input":["text","image"]},
- "open-mistral-7b": {"input":["text"]},
- "open-mistral-nemo": {"input":["text"]},
- "open-mixtral-8x22b": {"input":["text"]},
- "open-mixtral-8x7b": {"input":["text"]},
- "pixtral-12b": {"input":["text","image"]},
- "pixtral-large-latest": {"input":["text","image"]},
+ "codestral-latest": {"input":["text"],"context":256000},
+ "devstral-2512": {"input":["text"],"context":262144},
+ "devstral-latest": {"input":["text"],"context":262144},
+ "devstral-medium-2507": {"input":["text"],"context":128000},
+ "devstral-medium-latest": {"input":["text"],"context":262144},
+ "devstral-small-2505": {"input":["text"],"context":128000},
+ "devstral-small-2507": {"input":["text"],"context":128000},
+ "labs-devstral-small-2512": {"input":["text","image"],"context":256000},
+ "magistral-medium-latest": {"reasoning":true,"input":["text"],"context":128000},
+ "magistral-small": {"reasoning":true,"input":["text"],"context":128000},
+ "ministral-3b-latest": {"input":["text"],"context":128000},
+ "ministral-8b-latest": {"input":["text"],"context":128000},
+ "mistral-embed": {"input":["text"],"context":8000},
+ "mistral-large-2411": {"input":["text"],"context":131072},
+ "mistral-large-2512": {"input":["text","image"],"context":262144},
+ "mistral-large-latest": {"input":["text","image"],"context":262144},
+ "mistral-medium-2505": {"input":["text","image"],"context":131072},
+ "mistral-medium-2508": {"input":["text","image"],"context":262144},
+ "mistral-medium-2604": {"reasoning":true,"efforts":["none","high"],"input":["text","image"],"context":262144},
+ "mistral-medium-latest": {"reasoning":true,"efforts":["none","high"],"input":["text","image"],"context":262144},
+ "mistral-nemo": {"input":["text"],"context":128000},
+ "mistral-small-2506": {"input":["text","image"],"context":128000},
+ "mistral-small-2603": {"reasoning":true,"efforts":["none","high"],"input":["text","image"],"context":256000},
+ "mistral-small-latest": {"reasoning":true,"efforts":["none","high"],"input":["text","image"],"context":256000},
+ "open-mistral-7b": {"input":["text"],"context":8000},
+ "open-mistral-nemo": {"input":["text"],"context":128000},
+ "open-mixtral-8x22b": {"input":["text"],"context":64000},
+ "open-mixtral-8x7b": {"input":["text"],"context":32000},
+ "pixtral-12b": {"input":["text","image"],"context":128000},
+ "pixtral-large-latest": {"input":["text","image"],"context":128000},
"voxtral-mini-latest": {"input":["audio"]},
"voxtral-mini-tts-latest": {"input":["text"]},
- "voxtral-small-latest": {"input":["text","audio"]},
- "zai-glm-5-2": {"reasoning":true,"efforts":["high","max"],"input":["text"]},
+ "voxtral-small-latest": {"input":["text","audio"],"context":32000},
+ "zai-glm-5-2": {"reasoning":true,"efforts":["high","max"],"input":["text"],"context":1000000},
},
"deepseek": {
- "deepseek-flash": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image"]},
- "deepseek-v4-flash": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image"]},
- "deepseek-v4-flash-vision-exp": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image"]},
- "deepseek-v4-pro": {"reasoning":true,"efforts":["high","max"],"toggle":true,"input":["text"]},
+ "deepseek-flash": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image"],"context":1000000},
+ "deepseek-v4-flash": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image"],"context":1000000},
+ "deepseek-v4-flash-vision-exp": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image"],"context":1000000},
+ "deepseek-v4-pro": {"reasoning":true,"efforts":["high","max"],"toggle":true,"input":["text"],"context":1000000},
},
"kimi": {
- "kimi-k2.6": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "kimi-k2.7-code": {"reasoning":true,"input":["text","image","video"]},
- "kimi-k2.7-code-highspeed": {"reasoning":true,"input":["text","image","video"]},
- "kimi-k3": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image","video"]},
+ "kimi-k2.6": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":262144},
+ "kimi-k2.7-code": {"reasoning":true,"input":["text","image","video"],"context":262144},
+ "kimi-k2.7-code-highspeed": {"reasoning":true,"input":["text","image","video"],"context":262144},
+ "kimi-k3": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image","video"],"context":1048576},
},
"cohere": {
- "c4ai-aya-expanse-32b": {"input":["text"]},
- "c4ai-aya-expanse-8b": {"input":["text"]},
- "c4ai-aya-vision-32b": {"input":["text","image"]},
- "c4ai-aya-vision-8b": {"input":["text","image"]},
- "command-a-03-2025": {"input":["text"]},
- "command-a-plus-05-2026": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "command-a-reasoning-08-2025": {"reasoning":true,"toggle":true,"input":["text"]},
- "command-a-translate-08-2025": {"input":["text"]},
- "command-a-vision-07-2025": {"input":["text","image"]},
- "command-r-08-2024": {"input":["text"]},
- "command-r-plus-08-2024": {"input":["text"]},
- "command-r7b-12-2024": {"input":["text"]},
- "command-r7b-arabic-02-2025": {"input":["text"]},
- "north-mini-code-1-0": {"reasoning":true,"efforts":["none","high"],"input":["text"]},
+ "c4ai-aya-expanse-32b": {"input":["text"],"context":128000},
+ "c4ai-aya-expanse-8b": {"input":["text"],"context":8000},
+ "c4ai-aya-vision-32b": {"input":["text","image"],"context":16000},
+ "c4ai-aya-vision-8b": {"input":["text","image"],"context":16000},
+ "command-a-03-2025": {"input":["text"],"context":256000},
+ "command-a-plus-05-2026": {"reasoning":true,"toggle":true,"input":["text","image"],"context":128000},
+ "command-a-reasoning-08-2025": {"reasoning":true,"toggle":true,"input":["text"],"context":256000},
+ "command-a-translate-08-2025": {"input":["text"],"context":8000},
+ "command-a-vision-07-2025": {"input":["text","image"],"context":128000},
+ "command-r-08-2024": {"input":["text"],"context":128000},
+ "command-r-plus-08-2024": {"input":["text"],"context":128000},
+ "command-r7b-12-2024": {"input":["text"],"context":128000},
+ "command-r7b-arabic-02-2025": {"input":["text"],"context":128000},
+ "north-mini-code-1-0": {"reasoning":true,"efforts":["none","high"],"input":["text"],"context":256000},
},
"huggingface": {
- "deepseek-ai/deepseek-r1": {"reasoning":true,"input":["text"]},
- "deepseek-ai/deepseek-r1-0528": {"reasoning":true,"input":["text"]},
- "deepseek-ai/deepseek-v3": {"input":["text"]},
- "deepseek-ai/deepseek-v3-0324": {"input":["text"]},
- "deepseek-ai/deepseek-v3.1": {"reasoning":true,"input":["text"]},
- "deepseek-ai/deepseek-v3.2": {"reasoning":true,"input":["text"]},
- "deepseek-ai/deepseek-v4-flash": {"reasoning":true,"input":["text"]},
- "deepseek-ai/deepseek-v4-flash-0731": {"reasoning":true,"efforts":["high","max"],"input":["text"]},
- "deepseek-ai/deepseek-v4-flash-vision-exp": {"reasoning":true,"efforts":["low","high","max"],"input":["text","image"]},
- "deepseek-ai/deepseek-v4-pro": {"reasoning":true,"efforts":["high"],"input":["text"]},
- "deepseek-ai/deepseek-v4-pro-0813": {"reasoning":true,"efforts":["low","high","max"],"input":["text"]},
- "deepseek-ai/deepseek-v4.1-flash": {"reasoning":true,"efforts":["none","low","high","xhigh","max"],"input":["text","image"]},
- "google/gemma-3-12b-it": {"input":["text","image"]},
- "google/gemma-3-27b-it": {"input":["text","image"]},
- "google/gemma-3-4b-it": {"input":["text","image"]},
- "google/gemma-4-26b-a4b-it": {"reasoning":true,"input":["text","image"]},
- "google/gemma-4-31b-it": {"reasoning":true,"input":["text","image"]},
- "meta-llama/llama-3.1-8b-instruct": {"input":["text"]},
- "meta-llama/llama-3.3-70b-instruct": {"input":["text"]},
- "minimaxai/minimax-m2": {"reasoning":true,"input":["text"]},
- "minimaxai/minimax-m2.1": {"reasoning":true,"input":["text"]},
- "minimaxai/minimax-m2.5": {"reasoning":true,"input":["text"]},
- "minimaxai/minimax-m2.7": {"reasoning":true,"input":["text"]},
- "minimaxai/minimax-m3": {"reasoning":true,"input":["text","image"]},
- "moonshotai/kimi-k2-instruct": {"input":["text"]},
- "moonshotai/kimi-k2-instruct-0905": {"input":["text"]},
- "moonshotai/kimi-k2-thinking": {"reasoning":true,"input":["text"]},
- "moonshotai/kimi-k2.5": {"reasoning":true,"input":["text","image","video"]},
- "moonshotai/kimi-k2.6": {"reasoning":true,"input":["text","image","video"]},
- "moonshotai/kimi-k2.7-code": {"reasoning":true,"input":["text","image"]},
- "moonshotai/kimi-k3": {"reasoning":true,"efforts":["low","high","max"],"input":["text","image"]},
- "openai/gpt-oss-120b": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"]},
- "openai/gpt-oss-20b": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"]},
- "qwen/qwen2.5-coder-32b-instruct": {"input":["text"]},
- "qwen/qwen3-235b-a22b": {"reasoning":true,"input":["text"]},
- "qwen/qwen3-235b-a22b-instruct-2507": {"input":["text"]},
- "qwen/qwen3-235b-a22b-thinking-2507": {"reasoning":true,"input":["text"]},
- "qwen/qwen3-30b-a3b": {"reasoning":true,"input":["text"]},
- "qwen/qwen3-32b": {"reasoning":true,"input":["text"]},
- "qwen/qwen3-coder-30b-a3b-instruct": {"input":["text"]},
- "qwen/qwen3-coder-480b-a35b-instruct": {"input":["text"]},
- "qwen/qwen3-coder-next": {"input":["text"]},
- "qwen/qwen3-embedding-4b": {"input":["text"]},
- "qwen/qwen3-embedding-8b": {"input":["text"]},
- "qwen/qwen3-next-80b-a3b-instruct": {"input":["text"]},
- "qwen/qwen3-next-80b-a3b-thinking": {"input":["text"]},
- "qwen/qwen3-vl-235b-a22b-instruct": {"input":["text","image"]},
- "qwen/qwen3-vl-235b-a22b-thinking": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image"]},
- "qwen/qwen3.5-122b-a10b": {"reasoning":true,"input":["text","image"]},
- "qwen/qwen3.5-27b": {"reasoning":true,"input":["text","image"]},
- "qwen/qwen3.5-35b-a3b": {"reasoning":true,"input":["text","image"]},
- "qwen/qwen3.5-397b-a17b": {"reasoning":true,"efforts":["none","low","medium","high"],"input":["text","image"]},
- "qwen/qwen3.5-9b": {"reasoning":true,"input":["text","image"]},
- "qwen/qwen3.6-27b": {"reasoning":true,"input":["text","image"]},
- "qwen/qwen3.6-35b-a3b": {"reasoning":true,"input":["text","image"]},
- "qwen/qwen3.8-2.4t-a95b": {"reasoning":true,"efforts":["low","medium","xhigh"],"input":["text"]},
- "qwen/qwen3.8-27b": {"reasoning":true,"efforts":["low","medium","xhigh"],"input":["text","image"]},
- "stepfun-ai/step-3.5-flash": {"reasoning":true,"input":["text"]},
- "stepfun-ai/step-3.7-flash": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video"]},
- "tencent/hy3": {"reasoning":true,"efforts":["none","low","high"],"input":["text"]},
- "thinkingmachines/inkling": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image"]},
- "thinkingmachines/inkling-small": {"reasoning":true,"input":["text","image"]},
- "xiaomimimo/mimo-v2-flash": {"reasoning":true,"input":["text"]},
- "xiaomimimo/mimo-v2.5": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"toggle":true,"input":["text"]},
- "xiaomimimo/mimo-v2.5-pro": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"toggle":true,"input":["text"]},
- "zai-org/glm-4.5": {"reasoning":true,"input":["text"]},
- "zai-org/glm-4.5-air": {"reasoning":true,"input":["text"]},
- "zai-org/glm-4.5v": {"reasoning":true,"input":["text","image"]},
- "zai-org/glm-4.6": {"reasoning":true,"input":["text"]},
- "zai-org/glm-4.6v-flash": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "zai-org/glm-4.7": {"reasoning":true,"input":["text"]},
- "zai-org/glm-4.7-flash": {"reasoning":true,"input":["text"]},
- "zai-org/glm-5": {"reasoning":true,"input":["text"]},
- "zai-org/glm-5.1": {"reasoning":true,"input":["text"]},
- "zai-org/glm-5.2": {"reasoning":true,"input":["text"]},
- "zai-org/glm-5.3": {"reasoning":true,"efforts":["low","high","max"],"input":["text"]},
- "zai-org/glm-5.3-flash": {"reasoning":true,"efforts":["low","high","max"],"input":["text","image"]},
+ "deepseek-ai/deepseek-r1": {"reasoning":true,"input":["text"],"context":64000},
+ "deepseek-ai/deepseek-r1-0528": {"reasoning":true,"input":["text"],"context":163840},
+ "deepseek-ai/deepseek-v3": {"input":["text"],"context":64000},
+ "deepseek-ai/deepseek-v3-0324": {"input":["text"],"context":163840},
+ "deepseek-ai/deepseek-v3.1": {"reasoning":true,"input":["text"],"context":131072},
+ "deepseek-ai/deepseek-v3.2": {"reasoning":true,"input":["text"],"context":163840},
+ "deepseek-ai/deepseek-v4-flash": {"reasoning":true,"input":["text"],"context":1048576},
+ "deepseek-ai/deepseek-v4-flash-0731": {"reasoning":true,"efforts":["high","max"],"input":["text"],"context":1048576},
+ "deepseek-ai/deepseek-v4-flash-vision-exp": {"reasoning":true,"efforts":["low","high","max"],"input":["text","image"],"context":1048576},
+ "deepseek-ai/deepseek-v4-pro": {"reasoning":true,"efforts":["high"],"input":["text"],"context":1048576},
+ "deepseek-ai/deepseek-v4-pro-0813": {"reasoning":true,"efforts":["low","high","max"],"input":["text"],"context":1000000},
+ "deepseek-ai/deepseek-v4.1-flash": {"reasoning":true,"efforts":["none","low","high","xhigh","max"],"input":["text","image"],"context":1048576},
+ "google/gemma-3-12b-it": {"input":["text","image"],"context":131072},
+ "google/gemma-3-27b-it": {"input":["text","image"],"context":131072},
+ "google/gemma-3-4b-it": {"input":["text","image"],"context":131072},
+ "google/gemma-4-26b-a4b-it": {"reasoning":true,"input":["text","image"],"context":262144},
+ "google/gemma-4-31b-it": {"reasoning":true,"input":["text","image"],"context":262144},
+ "meta-llama/llama-3.1-8b-instruct": {"input":["text"],"context":131072},
+ "meta-llama/llama-3.3-70b-instruct": {"input":["text"],"context":131072},
+ "minimaxai/minimax-m2": {"reasoning":true,"input":["text"],"context":204800},
+ "minimaxai/minimax-m2.1": {"reasoning":true,"input":["text"],"context":204800},
+ "minimaxai/minimax-m2.5": {"reasoning":true,"input":["text"],"context":204800},
+ "minimaxai/minimax-m2.7": {"reasoning":true,"input":["text"],"context":204800},
+ "minimaxai/minimax-m3": {"reasoning":true,"input":["text","image"],"context":524288},
+ "moonshotai/kimi-k2-instruct": {"input":["text"],"context":131072},
+ "moonshotai/kimi-k2-instruct-0905": {"input":["text"],"context":262144},
+ "moonshotai/kimi-k2-thinking": {"reasoning":true,"input":["text"],"context":262144},
+ "moonshotai/kimi-k2.5": {"reasoning":true,"input":["text","image","video"],"context":262144},
+ "moonshotai/kimi-k2.6": {"reasoning":true,"input":["text","image","video"],"context":262144},
+ "moonshotai/kimi-k2.7-code": {"reasoning":true,"input":["text","image"],"context":262144},
+ "moonshotai/kimi-k3": {"reasoning":true,"efforts":["low","high","max"],"input":["text","image"],"context":1000000},
+ "openai/gpt-oss-120b": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"],"context":131072},
+ "openai/gpt-oss-20b": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"],"context":131072},
+ "qwen/qwen2.5-coder-32b-instruct": {"input":["text"],"context":131072},
+ "qwen/qwen3-235b-a22b": {"reasoning":true,"input":["text"],"context":40960},
+ "qwen/qwen3-235b-a22b-instruct-2507": {"input":["text"],"context":262144},
+ "qwen/qwen3-235b-a22b-thinking-2507": {"reasoning":true,"input":["text"],"context":262144},
+ "qwen/qwen3-30b-a3b": {"reasoning":true,"input":["text"],"context":40960},
+ "qwen/qwen3-32b": {"reasoning":true,"input":["text"],"context":131072},
+ "qwen/qwen3-coder-30b-a3b-instruct": {"input":["text"],"context":262144},
+ "qwen/qwen3-coder-480b-a35b-instruct": {"input":["text"],"context":262144},
+ "qwen/qwen3-coder-next": {"input":["text"],"context":262144},
+ "qwen/qwen3-embedding-4b": {"input":["text"],"context":32000},
+ "qwen/qwen3-embedding-8b": {"input":["text"],"context":32000},
+ "qwen/qwen3-next-80b-a3b-instruct": {"input":["text"],"context":262144},
+ "qwen/qwen3-next-80b-a3b-thinking": {"input":["text"],"context":262144},
+ "qwen/qwen3-vl-235b-a22b-instruct": {"input":["text","image"],"context":131072},
+ "qwen/qwen3-vl-235b-a22b-thinking": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image"],"context":131072},
+ "qwen/qwen3.5-122b-a10b": {"reasoning":true,"input":["text","image"],"context":262144},
+ "qwen/qwen3.5-27b": {"reasoning":true,"input":["text","image"],"context":262144},
+ "qwen/qwen3.5-35b-a3b": {"reasoning":true,"input":["text","image"],"context":262144},
+ "qwen/qwen3.5-397b-a17b": {"reasoning":true,"efforts":["none","low","medium","high"],"input":["text","image"],"context":262144},
+ "qwen/qwen3.5-9b": {"reasoning":true,"input":["text","image"],"context":262144},
+ "qwen/qwen3.6-27b": {"reasoning":true,"input":["text","image"],"context":262144},
+ "qwen/qwen3.6-35b-a3b": {"reasoning":true,"input":["text","image"],"context":262144},
+ "qwen/qwen3.8-2.4t-a95b": {"reasoning":true,"efforts":["low","medium","xhigh"],"input":["text"],"context":262144},
+ "qwen/qwen3.8-27b": {"reasoning":true,"efforts":["low","medium","xhigh"],"input":["text","image"],"context":262144},
+ "stepfun-ai/step-3.5-flash": {"reasoning":true,"input":["text"],"context":262144},
+ "stepfun-ai/step-3.7-flash": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image","video"],"context":262144},
+ "tencent/hy3": {"reasoning":true,"efforts":["none","low","high"],"input":["text"],"context":262144},
+ "thinkingmachines/inkling": {"reasoning":true,"efforts":["low","medium","high"],"input":["text","image"],"context":1048576},
+ "thinkingmachines/inkling-small": {"reasoning":true,"input":["text","image"],"context":524288},
+ "xiaomimimo/mimo-v2-flash": {"reasoning":true,"input":["text"],"context":262144},
+ "xiaomimimo/mimo-v2.5": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"toggle":true,"input":["text"],"context":262144},
+ "xiaomimimo/mimo-v2.5-pro": {"reasoning":true,"efforts":["none","low","medium","high","xhigh"],"toggle":true,"input":["text"],"context":1048576},
+ "zai-org/glm-4.5": {"reasoning":true,"input":["text"],"context":131072},
+ "zai-org/glm-4.5-air": {"reasoning":true,"input":["text"],"context":131072},
+ "zai-org/glm-4.5v": {"reasoning":true,"input":["text","image"],"context":65536},
+ "zai-org/glm-4.6": {"reasoning":true,"input":["text"],"context":204800},
+ "zai-org/glm-4.6v-flash": {"reasoning":true,"toggle":true,"input":["text","image"],"context":131072},
+ "zai-org/glm-4.7": {"reasoning":true,"input":["text"],"context":204800},
+ "zai-org/glm-4.7-flash": {"reasoning":true,"input":["text"],"context":200000},
+ "zai-org/glm-5": {"reasoning":true,"input":["text"],"context":202752},
+ "zai-org/glm-5.1": {"reasoning":true,"input":["text"],"context":202752},
+ "zai-org/glm-5.2": {"reasoning":true,"input":["text"],"context":262144},
+ "zai-org/glm-5.3": {"reasoning":true,"efforts":["low","high","max"],"input":["text"],"context":1048576},
+ "zai-org/glm-5.3-flash": {"reasoning":true,"efforts":["low","high","max"],"input":["text","image"],"context":1048576},
},
"qwen": {
- "deepseek-v4-flash-0731": {"reasoning":true,"efforts":["high","max"],"toggle":true,"input":["text"]},
- "glm-5.2": {"reasoning":true,"efforts":["none","minimal","low","medium","high","xhigh","max"],"input":["text"]},
- "qvq-max": {"reasoning":true,"input":["text","image"]},
- "qwen-flash": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen-max": {"input":["text"]},
- "qwen-mt-plus": {"input":["text"]},
- "qwen-mt-turbo": {"input":["text"]},
- "qwen-omni-turbo": {"input":["text","image","audio","video"]},
- "qwen-omni-turbo-realtime": {"input":["text","image","audio"]},
- "qwen-plus": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen-plus-character-ja": {"input":["text"]},
- "qwen-turbo": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen-vl-max": {"input":["text","image"]},
- "qwen-vl-ocr": {"input":["text","image"]},
- "qwen-vl-plus": {"input":["text","image"]},
- "qwen2-5-14b-instruct": {"input":["text"]},
- "qwen2-5-32b-instruct": {"input":["text"]},
- "qwen2-5-72b-instruct": {"input":["text"]},
- "qwen2-5-7b-instruct": {"input":["text"]},
- "qwen2-5-omni-7b": {"input":["text","image","audio","video"]},
- "qwen2-5-vl-72b-instruct": {"input":["text","image"]},
- "qwen2-5-vl-7b-instruct": {"input":["text","image"]},
- "qwen3-14b": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen3-235b-a22b": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen3-32b": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen3-8b": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen3-asr-flash": {"input":["audio"]},
- "qwen3-coder-30b-a3b-instruct": {"input":["text"]},
- "qwen3-coder-480b-a35b-instruct": {"input":["text"]},
- "qwen3-coder-flash": {"input":["text"]},
- "qwen3-coder-plus": {"input":["text"]},
- "qwen3-livetranslate-flash-realtime": {"input":["text","image","audio","video"]},
- "qwen3-max": {"input":["text"]},
- "qwen3-next-80b-a3b-instruct": {"input":["text"]},
- "qwen3-next-80b-a3b-thinking": {"reasoning":true,"input":["text"]},
- "qwen3-omni-flash": {"reasoning":true,"toggle":true,"input":["text","image","audio","video"]},
- "qwen3-omni-flash-realtime": {"input":["text","image","audio","video"]},
- "qwen3-vl-235b-a22b": {"reasoning":true,"input":["text","image"]},
- "qwen3-vl-30b-a3b": {"reasoning":true,"input":["text","image"]},
- "qwen3-vl-plus": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "qwen3.5-122b-a10b": {"reasoning":true,"toggle":true,"input":["text","image","video","audio"]},
- "qwen3.5-27b": {"reasoning":true,"toggle":true,"input":["text","image","video","audio"]},
- "qwen3.5-35b-a3b": {"reasoning":true,"toggle":true,"input":["text","image","video","audio"]},
- "qwen3.5-397b-a17b": {"reasoning":true,"toggle":true,"input":["text","image","video","audio"]},
- "qwen3.5-plus": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen3.6-27b": {"reasoning":true,"toggle":true,"input":["text","image","video","audio"]},
- "qwen3.6-35b-a3b": {"reasoning":true,"toggle":true,"input":["text","image","video","audio"]},
- "qwen3.6-flash": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen3.6-max-preview": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen3.6-plus": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen3.7-max": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen3.7-plus": {"reasoning":true,"toggle":true,"input":["text","image","video"]},
- "qwen3.8-flash": {"reasoning":true,"efforts":["low","medium","xhigh"],"toggle":true,"input":["text","image","video"]},
- "qwen3.8-max": {"reasoning":true,"efforts":["low","medium","xhigh"],"toggle":true,"input":["text","image","video","pdf"]},
- "qwq-plus": {"reasoning":true,"input":["text"]},
+ "deepseek-v4-flash-0731": {"reasoning":true,"efforts":["high","max"],"toggle":true,"input":["text"],"context":1000000},
+ "glm-5.2": {"reasoning":true,"efforts":["none","minimal","low","medium","high","xhigh","max"],"input":["text"],"context":1000000},
+ "kimi-k3": {"reasoning":true,"efforts":["low","high","max"],"input":["text","image"],"context":1048576},
+ "qvq-max": {"reasoning":true,"input":["text","image"],"context":131072},
+ "qwen-flash": {"reasoning":true,"toggle":true,"input":["text"],"context":1000000},
+ "qwen-max": {"input":["text"],"context":32768},
+ "qwen-mt-plus": {"input":["text"],"context":16384},
+ "qwen-mt-turbo": {"input":["text"],"context":16384},
+ "qwen-omni-turbo": {"input":["text","image","audio","video"],"context":32768},
+ "qwen-omni-turbo-realtime": {"input":["text","image","audio"],"context":32768},
+ "qwen-plus": {"reasoning":true,"toggle":true,"input":["text"],"context":1000000},
+ "qwen-plus-character-ja": {"input":["text"],"context":8192},
+ "qwen-turbo": {"reasoning":true,"toggle":true,"input":["text"],"context":1000000},
+ "qwen-vl-max": {"input":["text","image"],"context":131072},
+ "qwen-vl-ocr": {"input":["text","image"],"context":34096},
+ "qwen-vl-plus": {"input":["text","image"],"context":131072},
+ "qwen2-5-14b-instruct": {"input":["text"],"context":131072},
+ "qwen2-5-32b-instruct": {"input":["text"],"context":131072},
+ "qwen2-5-72b-instruct": {"input":["text"],"context":131072},
+ "qwen2-5-7b-instruct": {"input":["text"],"context":131072},
+ "qwen2-5-omni-7b": {"input":["text","image","audio","video"],"context":32768},
+ "qwen2-5-vl-72b-instruct": {"input":["text","image"],"context":131072},
+ "qwen2-5-vl-7b-instruct": {"input":["text","image"],"context":131072},
+ "qwen3-14b": {"reasoning":true,"toggle":true,"input":["text"],"context":131072},
+ "qwen3-235b-a22b": {"reasoning":true,"toggle":true,"input":["text"],"context":131072},
+ "qwen3-32b": {"reasoning":true,"toggle":true,"input":["text"],"context":131072},
+ "qwen3-8b": {"reasoning":true,"toggle":true,"input":["text"],"context":131072},
+ "qwen3-asr-flash": {"input":["audio"],"context":53248},
+ "qwen3-coder-30b-a3b-instruct": {"input":["text"],"context":262144},
+ "qwen3-coder-480b-a35b-instruct": {"input":["text"],"context":262144},
+ "qwen3-coder-flash": {"input":["text"],"context":1000000},
+ "qwen3-coder-plus": {"input":["text"],"context":1048576},
+ "qwen3-livetranslate-flash-realtime": {"input":["text","image","audio","video"],"context":53248},
+ "qwen3-max": {"input":["text"],"context":262144},
+ "qwen3-next-80b-a3b-instruct": {"input":["text"],"context":131072},
+ "qwen3-next-80b-a3b-thinking": {"reasoning":true,"input":["text"],"context":131072},
+ "qwen3-omni-flash": {"reasoning":true,"toggle":true,"input":["text","image","audio","video"],"context":65536},
+ "qwen3-omni-flash-realtime": {"input":["text","image","audio","video"],"context":65536},
+ "qwen3-vl-235b-a22b": {"reasoning":true,"input":["text","image"],"context":131072},
+ "qwen3-vl-30b-a3b": {"reasoning":true,"input":["text","image"],"context":131072},
+ "qwen3-vl-plus": {"reasoning":true,"toggle":true,"input":["text","image"],"context":262144},
+ "qwen3.5-122b-a10b": {"reasoning":true,"toggle":true,"input":["text","image","video","audio"],"context":262144},
+ "qwen3.5-27b": {"reasoning":true,"toggle":true,"input":["text","image","video","audio"],"context":262144},
+ "qwen3.5-35b-a3b": {"reasoning":true,"toggle":true,"input":["text","image","video","audio"],"context":262144},
+ "qwen3.5-397b-a17b": {"reasoning":true,"toggle":true,"input":["text","image","video","audio"],"context":262144},
+ "qwen3.5-plus": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":1000000},
+ "qwen3.6-27b": {"reasoning":true,"toggle":true,"input":["text","image","video","audio"],"context":262144},
+ "qwen3.6-35b-a3b": {"reasoning":true,"toggle":true,"input":["text","image","video","audio"],"context":262144},
+ "qwen3.6-flash": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":1000000},
+ "qwen3.6-max-preview": {"reasoning":true,"toggle":true,"input":["text"],"context":262144},
+ "qwen3.6-plus": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":1000000},
+ "qwen3.7-max": {"reasoning":true,"toggle":true,"input":["text"],"context":1000000},
+ "qwen3.7-plus": {"reasoning":true,"toggle":true,"input":["text","image","video"],"context":1000000},
+ "qwen3.8-flash": {"reasoning":true,"efforts":["low","medium","xhigh"],"toggle":true,"input":["text","image","video"],"context":1000000},
+ "qwen3.8-max": {"reasoning":true,"efforts":["low","medium","xhigh"],"toggle":true,"input":["text","image","video","pdf"],"context":1000000},
+ "qwq-plus": {"reasoning":true,"input":["text"],"context":131072},
},
"ollama": {
- "deepseek-v4-flash": {"reasoning":true,"efforts":["high","max"],"toggle":true,"input":["text"]},
- "deepseek-v4-flash:0731": {"reasoning":true,"efforts":["high","max"],"toggle":true,"input":["text"]},
- "deepseek-v4-pro": {"reasoning":true,"efforts":["high","max"],"toggle":true,"input":["text"]},
- "deepseek-v4-pro:0813": {"reasoning":true,"efforts":["high","max"],"toggle":true,"input":["text"]},
- "deepseek-v4.1-flash": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image"]},
- "gemma4:31b": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "glm-5.1": {"reasoning":true,"toggle":true,"input":["text"]},
- "glm-5.2": {"reasoning":true,"efforts":["high","max"],"input":["text"]},
- "glm-5.3": {"reasoning":true,"efforts":["low","high","max"],"input":["text"]},
- "glm-5.3-flash": {"reasoning":true,"efforts":["low","high","max"],"input":["text","image","video","pdf"]},
- "gpt-oss:120b": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"]},
- "gpt-oss:20b": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"]},
- "kimi-k2.5": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "kimi-k2.6": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "kimi-k2.7-code": {"reasoning":true,"toggle":true,"input":["text","image"]},
- "kimi-k3": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image"]},
- "minimax-m2.5": {"reasoning":true,"input":["text"]},
- "minimax-m2.7": {"reasoning":true,"toggle":true,"input":["text"]},
- "minimax-m3": {"reasoning":true,"efforts":["low","medium","high","max"],"toggle":true,"input":["text","image","video"]},
- "mistral-large-3:675b": {"input":["text","image"]},
- "nemotron-3-nano:30b": {"reasoning":true,"toggle":true,"input":["text"]},
- "nemotron-3-super": {"reasoning":true,"toggle":true,"input":["text"]},
- "nemotron-3-ultra": {"reasoning":true,"toggle":true,"input":["text"]},
- "qwen3.5:397b": {"reasoning":true,"toggle":true,"input":["text","image"]},
+ "deepseek-v4-flash": {"reasoning":true,"efforts":["high","max"],"toggle":true,"input":["text"],"context":1048576},
+ "deepseek-v4-flash:0731": {"reasoning":true,"efforts":["high","max"],"toggle":true,"input":["text"],"context":1048576},
+ "deepseek-v4-pro": {"reasoning":true,"efforts":["high","max"],"toggle":true,"input":["text"],"context":1048576},
+ "deepseek-v4-pro:0813": {"reasoning":true,"efforts":["high","max"],"toggle":true,"input":["text"],"context":1048576},
+ "deepseek-v4.1-flash": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image"],"context":1048576},
+ "gemma4:31b": {"reasoning":true,"toggle":true,"input":["text","image"],"context":262144},
+ "glm-5.1": {"reasoning":true,"toggle":true,"input":["text"],"context":202752},
+ "glm-5.2": {"reasoning":true,"efforts":["high","max"],"input":["text"],"context":976000},
+ "glm-5.3": {"reasoning":true,"efforts":["low","high","max"],"input":["text"],"context":1048576},
+ "glm-5.3-flash": {"reasoning":true,"efforts":["low","high","max"],"input":["text","image","video","pdf"],"context":1000000},
+ "gpt-oss:120b": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"],"context":131072},
+ "gpt-oss:20b": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"],"context":131072},
+ "kimi-k2.5": {"reasoning":true,"toggle":true,"input":["text","image"],"context":262144},
+ "kimi-k2.6": {"reasoning":true,"toggle":true,"input":["text","image"],"context":262144},
+ "kimi-k2.7-code": {"reasoning":true,"toggle":true,"input":["text","image"],"context":262144},
+ "kimi-k3": {"reasoning":true,"efforts":["low","high","max"],"toggle":true,"input":["text","image"],"context":1048576},
+ "minimax-m2.5": {"reasoning":true,"input":["text"],"context":204800},
+ "minimax-m2.7": {"reasoning":true,"toggle":true,"input":["text"],"context":196608},
+ "minimax-m3": {"reasoning":true,"efforts":["low","medium","high","max"],"toggle":true,"input":["text","image","video"],"context":512000},
+ "mistral-large-3:675b": {"input":["text","image"],"context":262144},
+ "nemotron-3-nano:30b": {"reasoning":true,"toggle":true,"input":["text"],"context":1048576},
+ "nemotron-3-super": {"reasoning":true,"toggle":true,"input":["text"],"context":262144},
+ "nemotron-3-ultra": {"reasoning":true,"toggle":true,"input":["text"],"context":262144},
+ "qwen3.5:397b": {"reasoning":true,"toggle":true,"input":["text","image"],"context":262144},
},
"lmstudio": {
- "openai/gpt-oss-20b": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"]},
- "qwen/qwen3-30b-a3b-2507": {"input":["text"]},
- "qwen/qwen3-coder-30b": {"input":["text"]},
+ "openai/gpt-oss-20b": {"reasoning":true,"efforts":["low","medium","high"],"input":["text"],"context":131072},
+ "qwen/qwen3-30b-a3b-2507": {"input":["text"],"context":262144},
+ "qwen/qwen3-coder-30b": {"input":["text"],"context":262144},
},
};
diff --git a/studio/frontend/src/features/chat/model-catalog.ts b/studio/frontend/src/features/chat/model-catalog.ts
index 9fae61b5791..53468936027 100644
--- a/studio/frontend/src/features/chat/model-catalog.ts
+++ b/studio/frontend/src/features/chat/model-catalog.ts
@@ -35,6 +35,8 @@ export interface ModelCatalogEntry {
defaultEffort: ReasoningEffortLevel | null;
inputModalities: readonly string[] | null;
maxOutputTokens: number | null;
+ /** Total context window the provider publishes, or null when it publishes none. */
+ contextLength: number | null;
}
interface LiveCatalogRecord {
@@ -170,9 +172,21 @@ export function modelsDevCatalogFetchedAt(): number | null {
// Merged namespaces, rebuilt whenever a served catalog lands.
let mergedNamespaces: Map<string, Readonly<Record<string, ModelCatalogSnapshotEntry>>> | null = null;
+/** Provider types with no catalogue namespace of their own. A Codex subscription serves OpenAI's
+ * models under OpenAI's ids, and getExternalReasoningCapabilities already aliases it the same
+ * way; without this every Codex lookup resolved to null and read as "not published". */
+const CATALOG_NAMESPACE_ALIASES: Record<string, string> = {
+ openai_codex: "openai",
+};
+
+function catalogNamespace(providerType: string): string {
+ return CATALOG_NAMESPACE_ALIASES[providerType] ?? providerType;
+}
+
function snapshotNamespace(
providerType: string,
): Readonly<Record<string, ModelCatalogSnapshotEntry>> | undefined {
+ providerType = catalogNamespace(providerType);
hydrateModelsDev();
const served = modelsDev?.providers[providerType];
const bundled = MODEL_CATALOG_SNAPSHOT[providerType];
@@ -183,7 +197,14 @@ function snapshotNamespace(
if (!mergedNamespaces) mergedNamespaces = new Map();
const cached = mergedNamespaces.get(providerType);
if (cached) return cached;
- const merged = { ...bundled, ...served };
+ const merged: Record<string, ModelCatalogSnapshotEntry> = { ...bundled };
+ for (const [id, entry] of Object.entries(served)) {
+ // A served entry replaces the bundled one, except that it cannot take away a context window
+ // it has no field for: a backend older than that field, or its cached payload inside the
+ // day-long TTL, would otherwise report every model it covers as not publishing one.
+ const context = entry.context ?? bundled[id]?.context;
+ merged[id] = context == null ? entry : { ...entry, context };
+ }
mergedNamespaces.set(providerType, merged);
return merged;
}
@@ -203,6 +224,8 @@ function fromLiveModel(model: ProviderModelCapabilityInfo): ModelCatalogEntry {
typeof model.max_output_tokens === "number" && model.max_output_tokens > 0
? model.max_output_tokens
: null,
+ // The live response carries no context window; resolveModelCatalogEntry backfills it.
+ contextLength: null,
};
}
@@ -217,6 +240,7 @@ function fromSnapshotEntry(entry: ModelCatalogSnapshotEntry): ModelCatalogEntry
defaultEffort: null,
inputModalities: entry.input ?? null,
maxOutputTokens: null,
+ contextLength: entry.context ?? null,
};
}
@@ -285,14 +309,21 @@ export function resolveModelCatalogEntry(
const candidates = lookupCandidates(normalizedProvider, modelId);
if (candidates.length === 0) return null;
hydrateLiveCatalog();
+ const snapshot = snapshotNamespace(normalizedProvider);
const live = LIVE_CATALOG.get(normalizedProvider)?.models;
if (live) {
for (const candidate of candidates) {
const entry = live[candidate];
- if (entry) return entry;
+ if (!entry) continue;
+ // A live entry wins on what it states, but states no context window, so that one field
+ // falls through to the snapshot instead of reading as unpublished.
+ if (entry.contextLength != null || !snapshot) return entry;
+ const fallback = candidates
+ .map((id) => snapshot[id]?.context)
+ .find((context) => typeof context === "number" && context > 0);
+ return fallback == null ? entry : { ...entry, contextLength: fallback };
}
}
- const snapshot = snapshotNamespace(normalizedProvider);
if (!snapshot) return null;
for (const candidate of candidates) {
const entry = snapshot[candidate];
diff --git a/studio/frontend/src/features/chat/provider-capabilities.ts b/studio/frontend/src/features/chat/provider-capabilities.ts
index ae04c225c61..2dec9527046 100644
--- a/studio/frontend/src/features/chat/provider-capabilities.ts
+++ b/studio/frontend/src/features/chat/provider-capabilities.ts
@@ -93,6 +93,67 @@ export function clampReasoningEffortToLevels(
return effortLevels[0] ?? "low";
}
+/** Whether a level reaches the provider at all. The external request path sends reasoning_effort
+ * for that style alone; every other style carries a bare thinking on/off, so a level chosen for
+ * one of those is never sent. The default ladder is present either way, so the style is the only
+ * thing that can say so. */
+export function externalReasoningTakesEffort(
+ caps: ExternalReasoningCapabilities,
+): boolean {
+ return caps.supportsReasoning && caps.reasoningStyle === "reasoning_effort";
+}
+
+/** The effort an external model should run at, pin first.
+ *
+ * One resolver, because there are three callers and they must not disagree: a model switch, the
+ * normalization that reruns on reload and on every provider resync, and the picker's own
+ * per-model editor. The pin used to be read by the switch alone, so a reload or a resync put the
+ * provider default back over it.
+ *
+ * Pass `pinned: null` to resolve as if nothing were pinned, which is what clearing a pin needs. */
+export function resolveExternalReasoningEffort(opts: {
+ caps: ExternalReasoningCapabilities;
+ providerType: string | null | undefined;
+ /** The chat's level now: the clamp target, and what a model that cannot think keeps. */
+ current: ReasoningEffortLevel;
+ /** The level pinned on this model's picker row. Ignored when the model no longer offers it. */
+ pinned?: string | null;
+ /** Restore an existing preference without choosing a new model default. */
+ restore?: boolean;
+}): ReasoningEffortLevel {
+ const { caps, providerType, current, pinned } = opts;
+ const levels = caps.reasoningEffortLevels;
+ // A style that sends no level has nothing to resolve, and moving the chat's level for it would
+ // change what every other model runs at on the strength of a setting this one never sends.
+ if (!externalReasoningTakesEffort(caps)) return current;
+ // Set deliberately, for this model, so it outranks every default below.
+ if (pinned && levels.includes(pinned as ReasoningEffortLevel)) {
+ return pinned as ReasoningEffortLevel;
+ }
+ if (opts.restore) return clampReasoningEffortToLevels(current, levels);
+ if (caps.defaultEffort && levels.includes(caps.defaultEffort)) {
+ return caps.defaultEffort;
+ }
+ const clamped = clampReasoningEffortToLevels(current, levels);
+ // Anthropic gets the highest level, since Claude's adaptive thinking adjusts cost per turn;
+ // OpenAI gets "high"; everyone else "medium". Overridable via Think.
+ if (providerType === "anthropic") {
+ return levels.includes("xhigh")
+ ? "xhigh"
+ : levels.includes("high")
+ ? "high"
+ : clamped;
+ }
+ if (providerType === "openai") {
+ return levels.includes("high")
+ ? "high"
+ : levels.includes("medium")
+ ? "medium"
+ : clamped;
+ }
+ return levels.includes("medium") ? "medium" : clamped;
+}
+
/** Fallback cap for a model with no documented limit and no connection override. */
export const EXTERNAL_MAX_OUTPUT_TOKENS = 32768;
diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
index f3a3368b487..0d5c877fa07 100644
--- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
@@ -6,6 +6,11 @@ import {
mirrorHfTokenInto,
useHfTokenStore,
} from "@/features/hub/stores/hf-token-store";
+// eslint-disable-next-line no-restricted-imports -- Leaf module; the model-picker index pulls in chat.
+import {
+ pinnedReasoningEffort,
+ useModelReasoningEffortStore,
+} from "@/features/model-picker/components/model-selector/model-reasoning-effort";
import { loadedContextFields } from "@/features/model-picker/model-config/per-model-config";
import {
cachedPinnableGpuIndexKind,
@@ -41,7 +46,12 @@ import {
normalizeProjectAttachmentTarget,
type ProjectAttachmentTarget,
} from "../utils/project-attachment-target";
-import { getExternalMaxOutputTokens } from "../provider-capabilities";
+import {
+ type ExternalReasoningCapabilities,
+ externalReasoningTakesEffort,
+ getExternalMaxOutputTokens,
+ resolveExternalReasoningEffort,
+} from "../provider-capabilities";
import {
PERSISTED_INFERENCE_PARAM_KEYS,
REMEMBERED_INFERENCE_PARAM_KEYS,
@@ -866,8 +876,11 @@ function readThreadScopedSettings(
export function threadScopedOverride<K extends ThreadScopedSettingKey>(
key: K,
): ThreadScopedSettings[K] | undefined {
- // activeThreadScopedSettings only refreshes on the debounce, so for 400ms it holds pre-edit
- // values a load would revert and persist; prefer the store. A pending pin answers first.
+ // Effort is captured before a model pin can overwrite the live value.
+ if (key === "reasoningEffort" && activeThreadScopedSettings?.reasoningEffort !== undefined) {
+ return activeThreadScopedSettings.reasoningEffort as ThreadScopedSettings[K];
+ }
+ // Other pending edits still live in the store until the debounce.
if (
threadSettingsWriteThreadId !== null &&
threadSettingsWriteThreadId === threadScopedSettingsThreadId
@@ -917,6 +930,29 @@ function keepsStoredValueUnderConstraint(
);
}
+/** Whether a per-model pin currently owns the live `reasoningEffort`. The pin is written into the
+ * store so the composer shows it, but it belongs to the model, not to the chat. */
+function pinOwnsLiveReasoningEffort(state: ChatRuntimeStore): boolean {
+ return (
+ externalReasoningTakesEffort(state) &&
+ pinnedReasoningEffort(
+ state.params.checkpoint,
+ state.reasoningEffortLevels,
+ ) !== null
+ );
+}
+
+/** The chat's own effort while a pin hides it: what the chat is on record as running with, so a
+ * snapshot taken under a pin stores the chat's level and not the model's. */
+function reasoningEffortOnRecord(): ReasoningEffort | undefined {
+ return (
+ activeThreadScopedSettings?.reasoningEffort ??
+ globalThreadScopedDefaults?.reasoningEffort ??
+ effortDisplacedByPin ??
+ undefined
+ );
+}
+
// The pills chat-page clamps to the selected model's capabilities.
const CLAMPED_PILL_KEYS = [
"toolsEnabled",
@@ -1082,6 +1118,11 @@ function buildThreadScopedSnapshot(
) {
settings.reasoningEnabled = false;
}
+ // The thread record already includes explicit edits waiting on the debounce.
+ if (threadId === threadScopedSettingsThreadId && pinHoldsLiveEffort()) {
+ const onRecord = reasoningEffortOnRecord();
+ if (onRecord !== undefined) settings.reasoningEffort = onRecord;
+ }
// Same for every pill the model-selection pass clamps off without touching the snapshot:
// each pill's own capability rule is exactly when the user could not have done it.
if (threadId === threadScopedSettingsThreadId) {
@@ -1634,17 +1675,19 @@ function restoreDefaultsOverCommittedEdits(
/** What the user actually touched, read off the store, which still holds their edits. */
function heldThreadScopedChanges(
- held: { field: string }[],
+ held: { field: string; value?: unknown }[],
): ThreadScopedSettings {
const edited: Record<string, unknown> = {};
const live = useChatRuntimeStore.getState();
// Same reader the snapshot path uses: sampling keys sit under `params`, so a direct field
// read returns undefined and the sanitizer drops them.
for (const edit of held) {
- edited[edit.field] = readThreadScopedValue(
- live,
- edit.field as ThreadScopedSettingKey,
- );
+ edited[edit.field] = edit.field === "reasoningEffort" && edit.value !== undefined
+ ? edit.value
+ : readThreadScopedValue(
+ live,
+ edit.field as ThreadScopedSettingKey,
+ );
}
return sanitizeThreadScopedSettings(edited);
}
@@ -1714,6 +1757,12 @@ function captureThreadScopedEdit(
if (threadId === null) return false;
// Both ids: between a switch and its snapshot arriving the store still holds the old values.
if (threadId === threadScopedSettingsThreadId) {
+ if (field === "reasoningEffort" && value !== undefined) {
+ activeThreadScopedSettings = {
+ ...activeThreadScopedSettings,
+ reasoningEffort: value as ReasoningEffort,
+ };
+ }
explicitlyEditedThreadFields.add(field);
// Set by the user now, so the chat stores this rather than what it had before a constraint moved the same field.
constraintSuppressedThreadFields.delete(field);
@@ -2527,6 +2576,12 @@ type ChatRuntimeStore = {
enabled: boolean,
options?: { persist?: boolean },
) => void;
+ /** Write one model's remembered params directly, live or not. For the picker's per-model
+ * editor; elsewhere an edit belongs to the loaded model and setParams covers it. */
+ setRememberedParamsForModel: (
+ modelId: string,
+ patch: PersistedInferenceParams,
+ ) => void;
setLastOpenRouterChosenModel: (chosen: string | null) => void;
setReasoningStyle: (style: ReasoningStyle) => void;
setReasoningEffort: (effort: ReasoningEffort) => void;
@@ -2734,6 +2789,11 @@ const SCALAR_SETTING_KEYS = [
// Ids this browser holds a local answer for. Hydration keeps these and merges the rest, so a
// pre-hydration edit cannot drop other models.
const locallyRememberedModels = new Set<string>();
+// Per-model edits made before the settings response landed, held as PATCHES rather than rows.
+// The set above is overlaid wholesale, which is right for a row built from a hydrated entry but
+// not for one typed before there was an entry: that row names only the keys the user touched, so
+// replacing with it would drop the temperature, top-p or seed the response is about to bring.
+const modelParamEditsBeforeHydration = new Map<string, PersistedInferenceParams>();
/** Deferred, not module scope: per-model-params is in the @/features/chat import cycle, so
* naming PERSISTED_INFERENCE_PARAM_KEYS here reads a const in its TDZ and throws at import
* time (as watchedStorageKeys() avoids). Memoized, or the `+= 1` bumps stop accumulating. */
@@ -3228,6 +3288,12 @@ function getHydratedSettingsState(
hydrated[modelId] = local;
}
}
+ for (const [modelId, patch] of modelParamEditsBeforeHydration) {
+ hydrated[modelId] = { ...hydrated[modelId], ...patch };
+ // A complete row from here, so a later response takes the wholesale fence above.
+ locallyRememberedModels.add(modelId);
+ }
+ modelParamEditsBeforeHydration.clear();
// The entry arriving for this model predates the fenced edit, so lay the edit over it or the
// next defaults update replays the stale one.
if (checkpoint) {
@@ -3389,10 +3455,16 @@ function setScalarSettingVersion<K extends ScalarSettingKey>(
return;
}
const writeGlobal = () => {
+ if (key === "reasoningEffort" && globalThreadScopedDefaults !== null) {
+ globalThreadScopedDefaults = {
+ ...globalThreadScopedDefaults,
+ reasoningEffort: value as ReasoningEffort,
+ };
+ }
scalarSettingMutationVersions[key] += 1;
saveSettingsPatch({ [key]: value });
};
- if (captureThreadScopedEdit(key, writeGlobal)) return;
+ if (captureThreadScopedEdit(key, writeGlobal, value)) return;
writeGlobal();
}
@@ -3412,6 +3484,77 @@ function localQwenMigrationSettings(
};
}
+// The chat's own effort, kept when a per-model pin took its place in the live store so that
+// clearing the pin can put it back. A pin is applied with a plain setState and persists nothing,
+// and the thread snapshot is read off the live store, so nothing else records what it displaced.
+// Session only: after a reload the live level already IS the pin's and nothing preceded it here,
+// so clearing then resolves without it.
+let effortDisplacedByPin: ReasoningEffort | null = null;
+
+/** Called before a per-model pin overwrites the live effort. The first displacement wins, so
+ * pinning twice and then clearing returns to the chat's own level, not to the earlier pin. */
+export function noteEffortDisplacedByPin(current: ReasoningEffort): void {
+ // Nothing before hydration: the level on screen is still this store's own default, not the
+ // chat's, and hydration replaces it because a pin advances no mutation version. The pin sites
+ // rerun once it lands, and the first record is the one that sticks.
+ if (!useChatRuntimeStore.getState().settingsHydrated) return;
+ effortDisplacedByPin ??= current;
+}
+
+/** Whether the live effort is a pin's rather than the chat's own. A pin in force owns it outright:
+ * every path that applies one writes it, the snapshot apply holds the chat's level back behind it,
+ * and stating a level in the composer clears the pin, so the two cannot disagree. The record
+ * covers the pin that has since gone, whether cleared or withdrawn by a catalogue refresh, while
+ * the level in the store is still it. */
+export function pinHoldsLiveEffort(): boolean {
+ return (
+ pinOwnsLiveReasoningEffort(useChatRuntimeStore.getState()) ||
+ effortDisplacedByPin !== null
+ );
+}
+
+/** The level to resolve from when a pin is cleared, or null when neither source has one. The
+ * chat's own level first: the snapshot keeps it because buildThreadScopedSnapshot holds the pin
+ * back, so it survives a reload and a thread switch, which the in-memory record does not. */
+export function takeEffortDisplacedByPin(): ReasoningEffort | null {
+ const displaced = effortDisplacedByPin;
+ effortDisplacedByPin = null;
+ return (
+ activeThreadScopedSettings?.reasoningEffort ??
+ globalThreadScopedDefaults?.reasoningEffort ??
+ displaced
+ );
+}
+
+/** Apply a model pin without persisting it as the chat's preference. */
+export function reconcilePinnedReasoningEffort(opts: {
+ checkpoint: string;
+ caps: ExternalReasoningCapabilities;
+ providerType: string | null | undefined;
+}): void {
+ const state = useChatRuntimeStore.getState();
+ if (state.params.checkpoint !== opts.checkpoint) return;
+ const pinned = externalReasoningTakesEffort(opts.caps)
+ ? pinnedReasoningEffort(opts.checkpoint, opts.caps.reasoningEffortLevels)
+ : null;
+ if (!pinned && !pinHoldsLiveEffort()) return;
+ const next = resolveExternalReasoningEffort({
+ caps: opts.caps,
+ providerType: opts.providerType,
+ current: pinned
+ ? state.reasoningEffort
+ : (takeEffortDisplacedByPin() ?? state.reasoningEffort),
+ pinned,
+ restore: pinned === null,
+ });
+ if (pinned) noteEffortDisplacedByPin(state.reasoningEffort);
+ if (next === state.reasoningEffort) return;
+ useChatRuntimeStore.setState((live) => ({
+ reasoningEffort: next,
+ queuedSettingsEpoch: live.queuedSettingsEpoch + 1,
+ }));
+}
+
function installationReasoningEnabled(state: ChatRuntimeStore): boolean {
return threadScopedOverride("reasoningEnabled") !== undefined
? (globalThreadScopedDefaults?.reasoningEnabled ?? state.reasoningEnabled)
@@ -4713,8 +4856,14 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
// Edits made while this chat's snapshot was in flight: keep them and store them on the chat,
// or the read would silently undo a click the user already saw.
const heldFields = new Set<string>();
+ let heldEffort: ReasoningEffort | undefined;
if (threadId !== null && threadId === pendingPairingThreadId) {
- for (const edit of heldThreadScopedEdits) heldFields.add(edit.field);
+ for (const edit of heldThreadScopedEdits) {
+ heldFields.add(edit.field);
+ if (edit.field === "reasoningEffort" && edit.value !== undefined) {
+ heldEffort = edit.value as ReasoningEffort;
+ }
+ }
heldThreadScopedEdits = [];
pendingPairingThreadId = null;
// The window is over, so the next visit samples afresh rather than reusing this round's.
@@ -4761,6 +4910,14 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
}
hydratedDefaultsByHeldField.delete(field);
}
+ // Same reason as the snapshot: a pin applied before the first chat opened is in the live
+ // effort, and capturing it here would make one model's level the default every
+ // snapshot-less chat follows.
+ if (pinOwnsLiveReasoningEffort(state) && !heldFields.has("reasoningEffort")) {
+ const onRecord = reasoningEffortOnRecord();
+ if (onRecord === undefined) delete captured.reasoningEffort;
+ else captured.reasoningEffort = onRecord;
+ }
globalThreadScopedDefaults = captured as ThreadScopedSettings;
}
threadScopedSettingsThreadId = threadId;
@@ -4779,7 +4936,14 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
for (const key of THREAD_SCOPED_SETTING_KEYS) {
// The user set this one while the read was in flight, so it wins over what came back.
if (heldFields.has(key)) {
- applied[key] = readThreadScopedValue(state, key);
+ if (key === "reasoningEffort" && heldEffort !== undefined) {
+ applied[key] = heldEffort;
+ activeThreadScopedSettings = { ...activeThreadScopedSettings, reasoningEffort: heldEffort };
+ if (pinOwnsLiveReasoningEffort(state)) effortDisplacedByPin = heldEffort;
+ else nextState.reasoningEffort = heldEffort;
+ } else {
+ applied[key] = readThreadScopedValue(state, key);
+ }
continue;
}
// Full access was accepted through a warning dialog, so a switch must not drop it. The chat
@@ -4808,6 +4972,17 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
);
if (value === undefined) continue;
applied[key] = value;
+ // The pin outranks the chat for as long as its model is selected, so the incoming level
+ // stays the chat's on record, recorded just above, but must not land in the store over
+ // the pin. Same shape as the "full" case: recorded, not applied.
+ if (key === "reasoningEffort" && pinOwnsLiveReasoningEffort(state)) {
+ // This chat's own level, which the pin is now sitting on top of, so leaving the model
+ // hands it back. Assigned rather than noted: it replaces the last chat's, and the pin
+ // may have been in the store since before this one opened, displacing nothing then.
+ // The value is one the enum sanitizer already passed.
+ effortDisplacedByPin = value as ReasoningEffort;
+ continue;
+ }
if (isSameThreadScopedValue(value, readThreadScopedValue(state, key))) {
continue;
}
@@ -4870,6 +5045,9 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
// Mirror setCheckpoint's persistence: dropping the checkpoint must also clear any stored external selection.
saveLastExternalCheckpoint(null);
saveBool(CHAT_DEEP_RESEARCH_ENABLED_KEY, false);
+ const restoredEffort = pinHoldsLiveEffort()
+ ? takeEffortDisplacedByPin()
+ : null;
return set((state) => ({
queuedSettingsEpoch: state.queuedSettingsEpoch + 1,
// An unload leaves the model the same way a switch does, so record what it was running with.
@@ -4897,6 +5075,7 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
reasoningAlwaysOn: false,
reasoningEnabled: true,
reasoningStyle: "enable_thinking",
+ reasoningEffort: restoredEffort ?? state.reasoningEffort,
supportsReasoningOff: false,
reasoningEffortLevels: ["low", "medium", "high"],
supportsPreserveThinking: false,
@@ -5003,8 +5182,66 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
reasoningStyle,
queuedSettingsEpoch: state.queuedSettingsEpoch + 1,
})),
- setReasoningEffort: (reasoningEffort) =>
+ setRememberedParamsForModel: (modelId, patch) =>
set((state) => {
+ if (!modelId || !hasKeys(patch)) return state;
+ const merged = { ...state.paramsByModel[modelId], ...patch };
+ const next = { ...state.paramsByModel, [modelId]: merged };
+ // Only the moved keys: the server merges per key, so a full snapshot clobbers other tabs.
+ // Not gated on hydration, unlike a defaults write: this is a typed edit for one model that
+ // can only set the keys it names, and gating it dropped an edit made while the initial
+ // /api/chat/settings request was out.
+ saveSettingsPatch({ inferenceParamsByModel: { [modelId]: patch } });
+ // trackParamsByModel files nothing before hydration, so without this the response in flight
+ // rebuilds paramsByModel from the server and the edit is gone. Held as a patch: there is no
+ // entry to have built a row from yet.
+ if (!state.settingsHydrated) {
+ modelParamEditsBeforeHydration.set(modelId, {
+ ...modelParamEditsBeforeHydration.get(modelId),
+ ...patch,
+ });
+ }
+ // Editing the live model must land on the live params, not just on the next switch back.
+ // Through the restore a switch uses, since systemPrompt is one of the chat's own keys: the
+ // open chat outranks the model it is running, and writing past that would both send the
+ // model's prompt for the rest of the chat and let the next snapshot store it as the chat's.
+ // Live params only, as there: the row above keeps the model's own value.
+ const live = state.params.checkpoint === modelId;
+ const liveParams = live
+ ? restoreThreadScopedParams({ ...state.params, ...patch })
+ : null;
+ // Nothing left to apply once the chat has taken its keys back.
+ const liveChanged =
+ liveParams !== null &&
+ shouldAdvanceQueuedSettingsEpoch(state.params, liveParams);
+ // And those keys are fenced the way a slider edit fences its own, or a response in flight
+ // puts the global set back over them.
+ if (liveChanged) getChangedInferenceParams(liveParams, state.params);
+ return {
+ paramsByModel: trackParamsByModel(state, next, modelId) ?? next,
+ // The epoch as setParams advances it: a queued prompt or a paste still reading its file
+ // captured the old prompt and cap, and this is what tells it they are no longer current.
+ ...(liveChanged
+ ? {
+ params: liveParams,
+ queuedSettingsEpoch: state.queuedSettingsEpoch + 1,
+ }
+ : {}),
+ };
+ }),
+ setReasoningEffort: (reasoningEffort) => {
+ // Choosing another level removes the current model's override.
+ const checkpoint = useChatRuntimeStore.getState().params.checkpoint;
+ const { effortByModel, setModelReasoningEffort } =
+ useModelReasoningEffortStore.getState();
+ const pinned = checkpoint ? effortByModel[checkpoint] : undefined;
+ // Re-picking the pin's own level states nothing new, so it keeps the pin.
+ if (checkpoint && pinned !== undefined && pinned !== reasoningEffort) {
+ setModelReasoningEffort(checkpoint, null);
+ }
+ set((state) => {
+ // A retained pin must keep the preference it displaced.
+ if (pinned !== reasoningEffort) effortDisplacedByPin = null;
setScalarSettingVersion(
"reasoningEffort",
reasoningEffort,
@@ -5014,7 +5251,8 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
reasoningEffort,
queuedSettingsEpoch: state.queuedSettingsEpoch + 1,
};
- }),
+ });
+ },
setPreserveThinking: (preserveThinking) =>
set((state) => {
setScalarSettingVersion(
diff --git a/studio/frontend/src/features/model-picker/components/model-selector.tsx b/studio/frontend/src/features/model-picker/components/model-selector.tsx
index 66c039c3f33..17ea4cafc75 100644
--- a/studio/frontend/src/features/model-picker/components/model-selector.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-selector.tsx
@@ -12,6 +12,8 @@ import { TooltipProvider } from "@/components/ui/tooltip";
import { ApiProviderLogo } from "@/features/chat/api-provider-logo";
import type { HfTaskFilter } from "@/features/hub/hooks/use-hub-model-search";
+// eslint-disable-next-line no-restricted-imports -- The settings barrel imports this feature back.
+import { useSettingsDialogStore } from "@/features/settings/stores/settings-dialog-store";
import { useT } from "@/i18n";
import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
import { cn } from "@/lib/utils";
@@ -320,6 +322,7 @@ function ModelSelectorContent({
onEject,
onFoldersChange,
onBrowseHub,
+ onConfigureConnection,
onModelsChange,
deleteDisabled,
className,
@@ -348,6 +351,7 @@ function ModelSelectorContent({
onEject?: () => void;
onFoldersChange?: () => void;
onBrowseHub?: () => void;
+ onConfigureConnection?: (providerId: string) => void;
onModelsChange?: (deletedModel?: DeletedModelRef) => void;
deleteDisabled?: boolean;
className?: string;
@@ -610,6 +614,7 @@ function ModelSelectorContent({
resolveDownloadFootprint={resolveDownloadFootprint}
onFoldersChange={onFoldersChange}
onBrowseHub={onBrowseHub}
+ onConfigureConnection={onConfigureConnection}
onModelsChange={onModelsChange}
onConfigure={openConfigPage}
deleteDisabled={deleteDisabled}
@@ -802,6 +807,13 @@ export function ModelSelector({
void navigate({ to: "/hub", search: { tab: "discover" } });
}
+ // A Connected group's gear. What is configurable about a remote model lives on its connection,
+ // so open that form rather than ModelConfigPage's local load settings.
+ function handleConfigureConnection(providerId: string) {
+ setOpen(false);
+ useSettingsDialogStore.getState().openConnectionSettings(providerId);
+ }
+
return (
<Popover open={open} onOpenChange={setOpen}>
<ModelSelectorTrigger
@@ -841,6 +853,7 @@ export function ModelSelector({
onBrowseHub={
task && communityModelPolicy === "none" ? undefined : handleBrowseHub
}
+ onConfigureConnection={handleConfigureConnection}
onModelsChange={onModelsChange}
deleteDisabled={deleteDisabled}
className={contentClassName}
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/connected-model-info-dialog.tsx b/studio/frontend/src/features/model-picker/components/model-selector/connected-model-info-dialog.tsx
new file mode 100644
index 00000000000..24306265d7c
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/components/model-selector/connected-model-info-dialog.tsx
@@ -0,0 +1,167 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// What the app knows about one connected model. A local row answers "what is this" with its size,
+// quant and path; a connected row has none of those, so everything comes from the provider
+// registry and the connection's cached catalogue.
+//
+// Each figure is read through the same resolver the rest of the app uses for it, so this cannot
+// disagree with the composer's own chips.
+//
+// What the model is, not what it is set to: the gear beside this opens the three settings it
+// keeps of its own and shows what each holds, so printing them here was that dialog again.
+
+import {
+ Dialog,
+ DialogContent,
+ DialogHeader,
+ DialogTitle,
+} from "@/components/ui/dialog";
+import { modelCatalogVersion, subscribeModelCatalog } from "@/features/chat";
+// eslint-disable-next-line no-restricted-imports -- Avoid the chat barrel's React exports.
+import { resolveModelCatalogEntry } from "@/features/chat/model-catalog";
+// eslint-disable-next-line no-restricted-imports -- Avoid the chat barrel's React exports.
+import {
+ externalReasoningTakesEffort,
+ getExternalReasoningCapabilities,
+ getPublishedExternalMaxOutputTokens,
+} from "@/features/chat/provider-capabilities";
+import { type ReactNode, useSyncExternalStore } from "react";
+import { connectedModelMarks } from "./connected-model-meta";
+
+const MODALITY_LABELS: Record<string, string> = {
+ text: "Text",
+ image: "Images",
+ audio: "Audio",
+ video: "Video",
+ pdf: "PDF",
+};
+
+function tokens(count: number): string {
+ return `${count.toLocaleString()} tokens`;
+}
+
+function Field({ label, children }: { label: string; children: ReactNode }) {
+ return (
+ <div className="flex items-baseline justify-between gap-6 py-2">
+ <span className="shrink-0 text-xs text-muted-foreground">{label}</span>
+ <span className="min-w-0 text-right text-xs text-foreground">
+ {children}
+ </span>
+ </div>
+ );
+}
+
+function Unset({ children }: { children: string }) {
+ return <span className="text-muted-foreground">{children}</span>;
+}
+
+export function ConnectedModelInfoDialog({
+ open,
+ onOpenChange,
+ modelId,
+ displayName,
+ providerName,
+ providerType,
+ baseUrl,
+ isReasoningProvider,
+}: {
+ open: boolean;
+ onOpenChange: (open: boolean) => void;
+ /** The provider's own id, which is what every lookup here keys on. */
+ modelId: string;
+ displayName: string;
+ providerName: string;
+ providerType: string;
+ baseUrl?: string | null;
+ /** A vLLM connection flagged as serving a reasoning model: the only signal a self-host gives. */
+ isReasoningProvider?: boolean;
+}) {
+ // Every figure below is read from the catalogue, which can land after this renders.
+ useSyncExternalStore(subscribeModelCatalog, modelCatalogVersion);
+ const entry = resolveModelCatalogEntry(providerType, modelId);
+ const marks = connectedModelMarks({ providerType, modelId, baseUrl });
+ // The resolver behind the composer's Thinking chip, not the catalogue alone.
+ const reasoning = getExternalReasoningCapabilities(providerType, modelId, {
+ isReasoningProvider,
+ baseUrl,
+ });
+ // "none" is the off switch, not a rung. Only where a level is sent at all: the default ladder
+ // is there even for a style that carries a bare thinking on/off, and listing it read as a
+ // choice this model takes.
+ const takesEffort = externalReasoningTakesEffort(reasoning);
+ const effortLevels = takesEffort
+ ? reasoning.reasoningEffortLevels.filter((level) => level !== "none")
+ : [];
+ const publishedMaxOutput =
+ getPublishedExternalMaxOutputTokens(providerType, modelId) ??
+ entry?.maxOutputTokens ??
+ null;
+
+ const inputs = entry?.inputModalities?.length
+ ? entry.inputModalities
+ .map((modality) => MODALITY_LABELS[modality] ?? modality)
+ .join(" · ")
+ : marks.vision
+ ? "Text · Images"
+ : "Text";
+ const generates = [
+ marks.capabilities.imageGen ? "Images" : null,
+ marks.capabilities.videoGen ? "Video" : null,
+ ].filter(Boolean);
+
+ return (
+ <Dialog open={open} onOpenChange={onOpenChange}>
+ <DialogContent className="sm:max-w-xl">
+ <DialogHeader>
+ <DialogTitle>{displayName}</DialogTitle>
+ </DialogHeader>
+
+ <div className="-mt-2 divide-y divide-border/50">
+ <Field label="Model ID">
+ {/* break-all, not break-words: an id has no spaces to break at. */}
+ <span className="break-all font-mono">{modelId}</span>
+ </Field>
+ <Field label="Connection">{providerName}</Field>
+ {baseUrl ? (
+ <Field label="Endpoint">
+ <span className="break-all font-mono">{baseUrl}</span>
+ </Field>
+ ) : null}
+ <Field label="Accepts">{inputs}</Field>
+ {generates.length > 0 ? (
+ <Field label="Generates">{generates.join(" · ")}</Field>
+ ) : null}
+ <Field label="Context window">
+ {entry?.contextLength ? (
+ <span className="tabular-nums">
+ {tokens(entry.contextLength)}
+ </span>
+ ) : (
+ <Unset>Not published</Unset>
+ )}
+ </Field>
+ <Field label="Max output">
+ {publishedMaxOutput ? (
+ <span className="tabular-nums">{tokens(publishedMaxOutput)}</span>
+ ) : (
+ <Unset>Follows the connection</Unset>
+ )}
+ </Field>
+ <Field label="Reasoning">
+ {reasoning.supportsReasoning ? (
+ <>
+ {effortLevels.length > 0
+ ? effortLevels.join(" · ")
+ : "Supported"}
+ {reasoning.reasoningAlwaysOn ? " (always on)" : null}
+ </>
+ ) : (
+ <Unset>No</Unset>
+ )}
+ </Field>
+ </div>
+ </DialogContent>
+ </Dialog>
+ );
+}
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/connected-model-meta.ts b/studio/frontend/src/features/model-picker/components/model-selector/connected-model-meta.ts
new file mode 100644
index 00000000000..97bf4e9b602
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/components/model-selector/connected-model-meta.ts
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Every mark comes from what the request path already resolves (capability registry, provider
+// gates), never from the model's name: a mark the request cannot honour is a promise the picker
+// breaks. An On Device badge may guess, since it describes a repo you have not fetched yet.
+
+// eslint-disable-next-line no-restricted-imports -- Avoid the chat barrel's React exports.
+import { providerModelSupportsVision } from "@/features/chat/external-providers";
+// eslint-disable-next-line no-restricted-imports -- Avoid the chat barrel's React exports.
+import { providerSupportsBuiltinImageGeneration } from "@/features/chat/provider-capabilities";
+import type { ModelCapabilities } from "./model-capabilities";
+
+export interface ConnectedModelMarks {
+ /** Capability glyphs (audio / generates images / generates video). */
+ capabilities: ModelCapabilities;
+ /** Reads images: drawn as the row's Vision badge rather than a glyph. */
+ vision: boolean;
+}
+
+/** The badges one Connected row should draw.
+ *
+ * `modelId` is the provider's own id (`gemini-2.5-flash-image`), not the `external::` id the
+ * picker selects by. `baseUrl` is the connection's and only image generation reads it. */
+export function connectedModelMarks(opts: {
+ providerType: string | null | undefined;
+ modelId: string | null | undefined;
+ baseUrl?: string | null;
+}): ConnectedModelMarks {
+ const { providerType, modelId, baseUrl } = opts;
+ // null is unknown, drawn as no badge rather than as a promise.
+ const vision = providerModelSupportsVision(providerType, modelId) === true;
+ return {
+ capabilities: {
+ vision,
+ // CAPABILITY_BADGES draws no reasoning glyph, so it is not worth a second catalogue lookup.
+ reasoning: false,
+ // The audio adapter resolves the active model out of `models`, which holds loaded local
+ // models only, so `add` rejects every audio file under a connected selection.
+ audio: false,
+ // The helper alone, no name fallback: the same gate the adapter and the Images pill use.
+ imageGen: providerSupportsBuiltinImageGeneration(
+ providerType,
+ modelId,
+ baseUrl,
+ ),
+ // Nothing we connect to serves video generation through the chat route, so the name would
+ // be the only evidence and a glyph resting on it promises what selecting the row cannot do.
+ videoGen: false,
+ },
+ vision,
+ };
+}
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/connected-model-settings-dialog.tsx b/studio/frontend/src/features/model-picker/components/model-selector/connected-model-settings-dialog.tsx
new file mode 100644
index 00000000000..6bdea37e9ef
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/components/model-selector/connected-model-settings-dialog.tsx
@@ -0,0 +1,258 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// One connected model's own settings. The system prompt and output cap are not a new store: Chat
+// has remembered both per model for as long as "Remember settings per model" has been on, under
+// the same checkpoint id. This just makes that memory visible and editable from the row.
+//
+// Reasoning effort is the one new value here; the chat's own is a single global level.
+
+import { Button } from "@/components/ui/button";
+import {
+ Dialog,
+ DialogContent,
+ DialogDescription,
+ DialogFooter,
+ DialogHeader,
+ DialogTitle,
+} from "@/components/ui/dialog";
+import { Input } from "@/components/ui/input";
+import { Label } from "@/components/ui/label";
+import {
+ Select,
+ SelectContent,
+ SelectItem,
+ SelectTrigger,
+ SelectValue,
+} from "@/components/ui/select";
+import { Textarea } from "@/components/ui/textarea";
+import {
+ modelCatalogVersion,
+ reconcilePinnedReasoningEffort,
+ subscribeModelCatalog,
+ useChatRuntimeStore,
+} from "@/features/chat";
+// eslint-disable-next-line no-restricted-imports -- Avoid the chat barrel's React exports.
+import {
+ externalReasoningTakesEffort,
+ getExternalMaxOutputTokens,
+ getExternalMinOutputTokens,
+ getExternalReasoningCapabilities,
+} from "@/features/chat/provider-capabilities";
+import { useState, useSyncExternalStore } from "react";
+import { useModelReasoningEffortStore } from "./model-reasoning-effort";
+
+/** Follows the chat's own level. A Select cannot carry an empty value, so absence needs a name. */
+const FOLLOW_CHAT = "__follow_chat__";
+
+export function ConnectedModelSettingsDialog({
+ open,
+ onOpenChange,
+ checkpointId,
+ displayName,
+ modelId,
+ providerType,
+ baseUrl,
+ isReasoningProvider,
+ connectionMaxOutputTokens,
+}: {
+ open: boolean;
+ onOpenChange: (open: boolean) => void;
+ /** The `external::` id, which is what both memories key on. */
+ checkpointId: string;
+ displayName: string;
+ /** The provider's own id, for the catalogue lookup. */
+ modelId: string;
+ providerType: string;
+ baseUrl?: string | null;
+ /** A vLLM connection flagged as serving a reasoning model. */
+ isReasoningProvider?: boolean;
+ /** The connection's own output cap, which lowers the model's documented one. */
+ connectionMaxOutputTokens?: number | null;
+}) {
+ const remembered = useChatRuntimeStore(
+ (state) => state.paramsByModel[checkpointId],
+ );
+ const rememberParamsPerModel = useChatRuntimeStore(
+ (state) => state.rememberParamsPerModel,
+ );
+ const setRememberedParamsForModel = useChatRuntimeStore(
+ (state) => state.setRememberedParamsForModel,
+ );
+ // The chat's own cap, which is what a model with none of its own runs at.
+ const chatMaxTokens = useChatRuntimeStore((state) => state.params.maxTokens);
+ // Whether this is the model currently loaded, which decides whether an edit has to reach the
+ // live settings as well as the memory.
+ const isLiveModel = useChatRuntimeStore(
+ (state) => state.params.checkpoint === checkpointId,
+ );
+ const pinnedEffort = useModelReasoningEffortStore(
+ (state) => state.effortByModel[checkpointId],
+ );
+ const setModelReasoningEffort = useModelReasoningEffortStore(
+ (state) => state.setModelReasoningEffort,
+ );
+
+ // The resolver behind the composer's own Thinking control, not the catalogue alone, so the
+ // levels offered here are the ones the provider actually accepts. "none" is the off switch.
+ const reasoning = getExternalReasoningCapabilities(providerType, modelId, {
+ isReasoningProvider,
+ baseUrl,
+ });
+ // Offered only where a level is actually sent: the default low/medium/high ladder is present
+ // even for a model whose style carries a bare thinking on/off, so gating on supportsReasoning
+ // alone let a pin be set on Kimi that no request could ever carry.
+ const efforts = externalReasoningTakesEffort(reasoning)
+ ? reasoning.reasoningEffortLevels.filter((level) => level !== "none")
+ : [];
+
+ // An OpenRouter cap comes from the live catalogue, which can land after this renders, and the
+ // bounds below are read from it.
+ useSyncExternalStore(subscribeModelCatalog, modelCatalogVersion);
+ // The bounds the chat's own Max Tokens control uses and the adapter clamps every request to, so
+ // a value stored here cannot differ from the one the provider is actually sent.
+ const minCap = getExternalMinOutputTokens(providerType);
+ const maxCap = getExternalMaxOutputTokens(
+ providerType,
+ modelId,
+ connectionMaxOutputTokens,
+ );
+ const clampCap = (value: number) =>
+ Math.min(Math.max(value, minCap), maxCap);
+
+ // null is untouched, so an untouched field keeps reading the store and Save writes nothing for
+ // it. The dialog can open before chat settings hydrate, and a useState seeded from the empty
+ // store then held that emptiness after the response landed, so Save put it back over the prompt
+ // the server had. A pin another tab changes while this is open reads the same way.
+ const [promptDraft, setPromptDraft] = useState<string | null>(null);
+ const [capDraft, setCapDraft] = useState<string | null>(null);
+ const [effortDraft, setEffortDraft] = useState<string | null>(null);
+ const systemPrompt = promptDraft ?? remembered?.systemPrompt ?? "";
+ // Against the ladder as it stands now, since the catalogue subscription below can withdraw a
+ // level while this is open. A draft the model no longer offers reverts to what is stored and
+ // Save leaves the pin alone, rather than writing a level every resolver then refuses. A stored
+ // pin the ladder dropped reads as Follow chat for the same reason: that is what it now does.
+ const offered = (level: string): boolean =>
+ (efforts as readonly string[]).includes(level);
+ const liveEffortDraft =
+ effortDraft !== null && (effortDraft === FOLLOW_CHAT || offered(effortDraft))
+ ? effortDraft
+ : null;
+ const effort =
+ liveEffortDraft ??
+ (pinnedEffort && offered(pinnedEffort) ? pinnedEffort : FOLLOW_CHAT);
+ // Always a real number. A blank would have to mean "forget the cap", and nothing can express
+ // that: paramsByModel merges per key here and the settings row deep-merges on the server, so an
+ // omitted key keeps the old value and the clear would be dropped without saying so.
+ const maxTokens =
+ capDraft ?? String(clampCap(remembered?.maxTokens ?? chatMaxTokens));
+
+ function save() {
+ // Number, not parseInt: a number field accepts scientific notation, and parseInt stops at the
+ // "e", so 1e5 read as 1 and was saved clamped to the provider minimum. Rounded because the
+ // cap is a token count and the field admits a decimal.
+ const typedCap = Math.round(Number(maxTokens.trim()));
+ // Only what the user actually touched: this is a patch, and writing an untouched field would
+ // put whatever the dialog happened to be showing over the stored value.
+ setRememberedParamsForModel(checkpointId, {
+ ...(promptDraft !== null ? { systemPrompt: promptDraft } : {}),
+ // A blank or junk field leaves the cap alone rather than sending a zero as the limit.
+ ...(capDraft !== null && Number.isFinite(typedCap) && typedCap > 0
+ ? { maxTokens: clampCap(typedCap) }
+ : {}),
+ });
+ // Untouched writes nothing at all: another tab can change this pin while the dialog is open,
+ // and a save of an unrelated field would otherwise put the value this select opened with back
+ // over it. It would also reset the composer's own Think level for no reason.
+ if (liveEffortDraft !== null) {
+ const pinned = liveEffortDraft === FOLLOW_CHAT ? null : liveEffortDraft;
+ setModelReasoningEffort(checkpointId, pinned);
+ if (isLiveModel) {
+ reconcilePinnedReasoningEffort({
+ checkpoint: checkpointId,
+ caps: reasoning,
+ providerType,
+ });
+ }
+ }
+ onOpenChange(false);
+ }
+
+ return (
+ <Dialog open={open} onOpenChange={onOpenChange}>
+ <DialogContent className="sm:max-w-lg">
+ <DialogHeader>
+ <DialogTitle>{displayName} settings</DialogTitle>
+ <DialogDescription>
+ Model defaults. Chats with their own system prompt keep it.
+ </DialogDescription>
+ </DialogHeader>
+
+ {rememberParamsPerModel ? null : (
+ <p className="rounded-md border border-border/70 bg-muted/30 px-3 py-2 text-xs leading-relaxed text-muted-foreground">
+ "Remember settings per model" is off in Settings → Chat, so the
+ prompt and output cap below are stored but never restored. Reasoning
+ effort is kept separately and still applies.
+ </p>
+ )}
+
+ <div className="flex flex-col gap-4">
+ <div className="flex flex-col gap-1.5">
+ <Label htmlFor="connected-model-system-prompt">System prompt</Label>
+ <Textarea
+ id="connected-model-system-prompt"
+ value={systemPrompt}
+ onChange={(event) => setPromptDraft(event.target.value)}
+ rows={5}
+ />
+ </div>
+
+ <div className="flex items-center justify-between gap-4">
+ <Label htmlFor="connected-model-max-tokens">
+ Max output tokens
+ <span className="ml-1 font-normal text-muted-foreground">
+ ({minCap.toLocaleString()} to {maxCap.toLocaleString()})
+ </span>
+ </Label>
+ <Input
+ id="connected-model-max-tokens"
+ type="number"
+ min={minCap}
+ max={maxCap}
+ inputMode="numeric"
+ value={maxTokens}
+ onChange={(event) => setCapDraft(event.target.value)}
+ className="w-28"
+ />
+ </div>
+
+ {efforts.length > 0 ? (
+ <div className="flex items-center justify-between gap-4">
+ <Label htmlFor="connected-model-effort">Reasoning effort</Label>
+ <Select value={effort} onValueChange={setEffortDraft}>
+ <SelectTrigger id="connected-model-effort" className="w-40">
+ <SelectValue />
+ </SelectTrigger>
+ <SelectContent>
+ <SelectItem value={FOLLOW_CHAT}>Follow chat</SelectItem>
+ {efforts.map((level) => (
+ <SelectItem key={level} value={level}>
+ {level}
+ </SelectItem>
+ ))}
+ </SelectContent>
+ </Select>
+ </div>
+ ) : null}
+ </div>
+
+ <DialogFooter>
+ <Button variant="outline" onClick={() => onOpenChange(false)}>
+ Cancel
+ </Button>
+ <Button onClick={save}>Save</Button>
+ </DialogFooter>
+ </DialogContent>
+ </Dialog>
+ );
+}
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/model-load-settings-action.tsx b/studio/frontend/src/features/model-picker/components/model-selector/model-load-settings-action.tsx
index 9cd02a756d9..991d2920107 100644
--- a/studio/frontend/src/features/model-picker/components/model-selector/model-load-settings-action.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-selector/model-load-settings-action.tsx
@@ -14,10 +14,13 @@ export function ModelLoadSettingsAction({
ariaLabel,
onConfigure,
className,
+ tooltip = "Configure run settings before loading model",
}: {
ariaLabel: string;
onConfigure: () => void;
className?: string;
+ /** Hover copy. The default describes a local load, so Connected rows override it. */
+ tooltip?: string;
}) {
return (
<Tooltip delayDuration={0}>
@@ -45,7 +48,7 @@ export function ModelLoadSettingsAction({
</button>
</TooltipTrigger>
<TooltipContent side="top" className="tooltip-compact">
- Configure run settings before loading model
+ {tooltip}
</TooltipContent>
</Tooltip>
);
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/model-reasoning-effort.ts b/studio/frontend/src/features/model-picker/components/model-selector/model-reasoning-effort.ts
new file mode 100644
index 00000000000..5b9f38f7769
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/components/model-selector/model-reasoning-effort.ts
@@ -0,0 +1,101 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { create } from "zustand";
+
+const KEY = "unsloth_model_reasoning_effort";
+const MODEL_KEY_PREFIX = `${KEY}::`;
+type EffortStorage = Pick<Storage, "getItem" | "setItem" | "key" | "length">;
+
+function browserStorage(): EffortStorage | null {
+ try {
+ return typeof localStorage === "undefined" ? null : localStorage;
+ } catch {
+ return null;
+ }
+}
+
+function readEfforts(storage: EffortStorage | null): Record<string, string> {
+ const efforts: Record<string, string> = {};
+ if (!storage) return efforts;
+ try {
+ const legacy: unknown = JSON.parse(storage.getItem(KEY) ?? "{}");
+ if (legacy && typeof legacy === "object" && !Array.isArray(legacy)) {
+ Object.assign(
+ efforts,
+ Object.fromEntries(
+ Object.entries(legacy).filter(
+ ([, value]) => typeof value === "string" && value.length > 0,
+ ),
+ ),
+ );
+ }
+ } catch {
+ // A broken legacy map must not hide newer entries.
+ }
+ try {
+ for (let index = 0; index < storage.length; index += 1) {
+ const key = storage.key(index);
+ if (!key?.startsWith(MODEL_KEY_PREFIX)) continue;
+ const modelId = key.slice(MODEL_KEY_PREFIX.length);
+ const effort = storage.getItem(key);
+ if (!modelId || effort === null) continue;
+ if (effort) Object.assign(efforts, { [modelId]: effort });
+ else delete efforts[modelId];
+ }
+ } catch {
+ // Keep entries already read when storage becomes unavailable.
+ }
+ return efforts;
+}
+
+interface ModelReasoningEffortState {
+ effortByModel: Record<string, string>;
+ setModelReasoningEffort: (modelId: string, effort: string | null) => void;
+ syncFromStorage: () => void;
+}
+
+export function createModelReasoningEffortStore(storage: EffortStorage | null) {
+ return create<ModelReasoningEffortState>((set) => ({
+ effortByModel: readEfforts(storage),
+ setModelReasoningEffort: (modelId, effort) =>
+ set((state) => {
+ if (!modelId) return state;
+ const next = { ...state.effortByModel };
+ if (effort) Object.assign(next, { [modelId]: effort });
+ else delete next[modelId];
+ try {
+ // Independent writes cannot erase another model's pin.
+ // Empty entries prevent a cleared legacy pin from returning.
+ storage?.setItem(`${MODEL_KEY_PREFIX}${modelId}`, effort ?? "");
+ } catch {
+ // Keep the choice for this session.
+ }
+ return { effortByModel: next };
+ }),
+ syncFromStorage: () => set({ effortByModel: readEfforts(storage) }),
+ }));
+}
+
+const storage = browserStorage();
+export const useModelReasoningEffortStore = createModelReasoningEffortStore(storage);
+
+/** Return only a pin offered by the current model. */
+export function pinnedReasoningEffort(
+ modelId: string | null | undefined,
+ allowed: readonly string[] | null | undefined,
+): string | null {
+ if (!modelId) return null;
+ const effort = useModelReasoningEffortStore.getState().effortByModel[modelId];
+ if (!effort || (allowed && !allowed.includes(effort))) return null;
+ return effort;
+}
+
+if (typeof window !== "undefined") {
+ window.addEventListener("storage", (event) => {
+ if (event.storageArea && event.storageArea !== storage) return;
+ if (event.key === KEY || event.key === null || event.key.startsWith(MODEL_KEY_PREFIX)) {
+ useModelReasoningEffortStore.getState().syncFromStorage();
+ }
+ });
+}
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/model-row-menu.tsx b/studio/frontend/src/features/model-picker/components/model-selector/model-row-menu.tsx
index c72a1d33d64..a822aadb6f2 100644
--- a/studio/frontend/src/features/model-picker/components/model-selector/model-row-menu.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-selector/model-row-menu.tsx
@@ -40,6 +40,15 @@ import {
useState,
} from "react";
+/** A caller-supplied entry. Rendered under the pin and above cache/update, so delete stays last. */
+export interface ModelRowMenuItem {
+ key: string;
+ label: string;
+ icon: ReactNode;
+ onSelect: () => void;
+ disabled?: boolean;
+}
+
interface ModelRowMenuPin {
pinned: boolean;
/** Menu item labels, e.g. "Pin quant to the top" / "Unpin quant". */
@@ -83,6 +92,7 @@ export function ModelRowMenu({
iconClassName,
cachePath,
pin,
+ items,
update,
del,
}: {
@@ -92,6 +102,8 @@ export function ModelRowMenu({
/** Enables "Reveal in Finder" for cached repos. */
cachePath?: ModelRowMenuCachePath;
pin?: ModelRowMenuPin;
+ /** Extra entries for actions this menu has no shape of its own for. */
+ items?: readonly ModelRowMenuItem[];
update?: ModelRowMenuUpdate;
del?: ModelRowMenuDelete;
}) {
@@ -171,7 +183,7 @@ export function ModelRowMenu({
});
}, [cachePathRepoId, cachePathVariant]);
- if (!pin && !update && !del && !cachePath) return null;
+ if (!pin && !update && !del && !cachePath && !items?.length) return null;
return (
<>
@@ -215,6 +227,19 @@ export function ModelRowMenu({
<span>{pin.pinned ? pin.unpinLabel : pin.pinLabel}</span>
</DropdownMenuItem>
)}
+ {items?.map((item) => (
+ <DropdownMenuItem
+ key={item.key}
+ disabled={item.disabled}
+ onSelect={(e) => {
+ e.stopPropagation();
+ item.onSelect();
+ }}
+ >
+ {item.icon}
+ <span>{item.label}</span>
+ </DropdownMenuItem>
+ ))}
{cachePath && (
<DropdownMenuItem
onSelect={(e) => {
@@ -244,7 +269,9 @@ export function ModelRowMenu({
)}
{del && (
<>
- {(cachePath || pin || update) && <DropdownMenuSeparator />}
+ {(cachePath || pin || update || items?.length) && (
+ <DropdownMenuSeparator />
+ )}
<DropdownMenuItem
variant="destructive"
disabled={del.disabled}
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/pickers.tsx b/studio/frontend/src/features/model-picker/components/model-selector/pickers.tsx
index 060d9395749..b2e1fd04686 100644
--- a/studio/frontend/src/features/model-picker/components/model-selector/pickers.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-selector/pickers.tsx
@@ -26,7 +26,11 @@ import {
import {
chatModelLoaded,
isExternalModelId,
+ modelCatalogVersion,
+ parseExternalModelId,
+ subscribeModelCatalog,
useChatRuntimeStore,
+ useExternalProvidersStore,
} from "@/features/chat";
import type {
CachedGgufRepo,
@@ -77,6 +81,7 @@ import {
useModelMemory,
} from "@/hooks/use-model-memory";
import { useVramBudgetFraction } from "@/hooks/use-vram-budget-fraction";
+import { copyToClipboard } from "@/lib/copy-to-clipboard";
import { diffusionRouteSearch } from "@/lib/diffusion-route-search";
import { type GgufFitClass, requiredGgufMemoryGb } from "@/lib/gguf-fit";
import { extractParamLabel } from "@/lib/model-size";
@@ -89,6 +94,7 @@ import {
ArrowUpDownIcon,
AudioWave01Icon,
Cancel01Icon,
+ Copy01Icon,
DashboardCircleIcon,
Flag01Icon,
FlimSlateIcon,
@@ -99,6 +105,7 @@ import {
PinIcon,
RemoveCircleIcon,
Search01Icon,
+ Settings02Icon,
ViewIcon,
} from "@hugeicons/core-free-icons";
import { HugeiconsIcon } from "@hugeicons/react";
@@ -117,6 +124,7 @@ import {
useMemo,
useRef,
useState,
+ useSyncExternalStore,
} from "react";
import { useChatPickerInventory } from "../../inventory/use-chat-picker-inventory";
import {
@@ -137,6 +145,11 @@ import {
taskForMediaPick,
taskPickerRowMatches,
} from "./audio-picker-policy";
+import { ConnectedModelInfoDialog } from "./connected-model-info-dialog";
+import {
+ connectedModelMarks,
+} from "./connected-model-meta";
+import { ConnectedModelSettingsDialog } from "./connected-model-settings-dialog";
import { FolderBrowser } from "./folder-browser";
import {
type ModelCapabilities,
@@ -166,6 +179,7 @@ import {
loadedAt,
useModelLoadTimes,
} from "./model-usage";
+import { usePinnedConnectedModelsStore } from "./pinned-connected-models";
import {
makePinRank,
pinKey,
@@ -941,6 +955,85 @@ const ROW_ACTIONS_CLASS =
// affordance, and hiding it reads as the stalled download having no controls.
const ROW_ACTIONS_PINNED_CLASS = cn(ROW_ACTIONS_CLASS, "opacity-100");
+// Same box and glyph size as ModelLoadSettingsAction, so a heading's buttons sit in the same
+// column and hover the same size as the ones on the rows under it.
+const HEADING_ACTION_CLASS =
+ "flex size-5 shrink-0 items-center justify-center rounded-md text-muted-foreground/60 transition hover:bg-black/5 hover:text-foreground dark:hover:bg-white/10";
+
+/** A Connected group label. Wider than the On Device section labels, since nothing divides these
+ * groups but the gap, and foldable the same way. */
+function ConnectedGroupHeading({
+ icon,
+ label,
+ collapsed,
+ onToggle,
+ onConfigure,
+ configureLabel,
+}: {
+ icon: ReactNode;
+ label: string;
+ collapsed: boolean;
+ onToggle: () => void;
+ /** Open this group's connection. A connection belongs to the whole group, not to one row. */
+ onConfigure?: () => void;
+ configureLabel?: string;
+}) {
+ return (
+ // pt-3 and gap-1.5, the values ListLabel uses for the On Device sections, so the first
+ // heading sits at the same height on both tabs and its label at the same offset from its
+ // icon. pt-5 started this list 8px lower than that one.
+ <div className="group/heading flex items-center justify-between gap-1 px-2.5 pb-1 pt-3">
+ <span className="flex min-w-0 items-center gap-1.5 text-ui-10 font-semibold uppercase tracking-wider text-muted-foreground">
+ {icon}
+ <span className="min-w-0 truncate">{label}</span>
+ </span>
+ {/* -mr-2 takes the heading's own px-2.5 down to the rows' mr-0.5, and the buttons carry the
+ row gutter's box and overlap, so the chevron lands in the column each row's dots menu
+ does rather than 8px inside it. */}
+ <div className="-mr-2 flex shrink-0 items-center -space-x-0.5">
+ {onConfigure ? (
+ <Tooltip delayDuration={0}>
+ <TooltipTrigger asChild={true}>
+ <button
+ type="button"
+ onClick={onConfigure}
+ aria-label={configureLabel ?? "Connection settings"}
+ // Hidden until the heading is hovered, like a row's own gear. The chevron beside
+ // it stays: folding is what the heading is for.
+ className={cn(
+ HEADING_ACTION_CLASS,
+ "opacity-0 group-hover/heading:opacity-100 focus-visible:opacity-100 [@media(hover:none)]:opacity-100",
+ )}
+ >
+ <HugeiconsIcon
+ icon={Settings02Icon}
+ strokeWidth={1.75}
+ className="size-3"
+ />
+ </button>
+ </TooltipTrigger>
+ <TooltipContent side="bottom" className="tooltip-compact">
+ {configureLabel ?? "Connection settings"}
+ </TooltipContent>
+ </Tooltip>
+ ) : null}
+ <button
+ type="button"
+ onClick={onToggle}
+ aria-label={collapsed ? "Expand section" : "Collapse section"}
+ className={HEADING_ACTION_CLASS}
+ >
+ {collapsed ? (
+ <ChevronRightIcon className="size-3" />
+ ) : (
+ <ChevronDownIcon className="size-3" />
+ )}
+ </button>
+ </div>
+ </div>
+ );
+}
+
function ModelRow({
label,
meta,
@@ -1055,6 +1148,7 @@ function ModelRow({
label: parsed.formats.map((f) => f.label).join(" · "),
}
: null;
+ const leading = formatDot ? <FormatTag {...formatDot} /> : null;
// Only the selected row charts itself: a meter under every row turns a list you scan into a wall of charts.
const memorySegments = useModelMemory(selected ? memory : undefined, gpuGb);
@@ -1094,18 +1188,16 @@ function ModelRow({
>
<span className="flex min-w-0 flex-1 items-baseline">
{/* Fixed slot, so names start on one line with or without a dot. */}
- {aligned ? (
+ {/* Aligned lists hold the slot open either way, so names start on one line; an
+ unaligned row only takes the space when it has something for it. */}
+ {aligned || leading ? (
<span
className={cn(
"mr-1 flex shrink-0 items-center self-center",
META_COLUMN.format,
)}
>
- {formatDot ? <FormatTag {...formatDot} /> : null}
- </span>
- ) : formatDot ? (
- <span className="mr-1 flex shrink-0 items-center self-center">
- <FormatTag {...formatDot} />
+ {leading}
</span>
) : null}
{showOwner ? (
@@ -1294,9 +1386,13 @@ function ModelRow({
return (
<Tooltip delayDuration={700}>
<TooltipTrigger asChild={true}>{content}</TooltipTrigger>
+ {/* Right, not left: this panel is docked to the model button at the window's left edge,
+ so a row's own left edge is ~30px in and a tooltip opening that way ran off screen.
+ Narrower with it, and breaking on words rather than anywhere: max-w-xs let a long
+ connection name stretch the box to 320px, most of it slack beside the model id. */}
<TooltipContent
- side="left"
- className="tooltip-compact max-w-xs break-all"
+ side="right"
+ className="tooltip-compact max-w-[15rem] break-words"
>
{tooltipBody}
</TooltipContent>
@@ -2467,6 +2563,32 @@ const FORMAT_FILTER_OPTIONS: HubOption<FormatFilter>[] = (
};
});
+// Connected has no size, download date or load history, so it sorts by what a hosted catalogue
+// does offer: the connection, or the name.
+type ConnectedSortKey = "provider" | "name";
+
+const CONNECTED_SORT_OPTIONS: HubOption<ConnectedSortKey>[] = [
+ { value: "provider", label: "Connection" },
+ { value: "name", label: "Name" },
+];
+
+// Modality filter for Connected, mirroring the On Device format filter. The keys are the marks a
+// connected row can draw, so a filter cannot ask for a badge the list has no way to show.
+// No Audio: connectedModelMarks never marks a connected model as taking audio, since the
+// attachment adapter refuses it under an external selection, so the option could only ever
+// return an empty list.
+type ConnectedModalityFilter = "all" | "vision" | "imageGen";
+
+const CONNECTED_MODALITY_LABELS: Record<ConnectedModalityFilter, string> = {
+ all: "All",
+ vision: "Vision",
+ imageGen: "Image gen",
+};
+
+const CONNECTED_MODALITY_OPTIONS: HubOption<ConnectedModalityFilter>[] = (
+ Object.keys(CONNECTED_MODALITY_LABELS) as ConnectedModalityFilter[]
+).map((value) => ({ value, label: CONNECTED_MODALITY_LABELS[value] }));
+
/** Sort cached repos: by last-loaded, download date, size desc, or name. */
function sortCachedRepos<
T extends { repo_id: string; size_bytes: number; last_modified?: number },
@@ -2592,6 +2714,7 @@ export function HubModelPicker({
resolveDownloadFootprint,
onFoldersChange,
onBrowseHub,
+ onConfigureConnection,
onModelsChange,
onConfigure,
deleteDisabled = false,
@@ -2616,6 +2739,9 @@ export function HubModelPicker({
resolveDownloadFootprint?: ModelDownloadFootprintResolver;
onFoldersChange?: () => void;
onBrowseHub?: () => void;
+ /** Open one connection's own settings, for the gear on a Connected row: a model reached over
+ * the wire has no local run settings, so what is configurable lives on its connection. */
+ onConfigureConnection?: (providerId: string) => void;
onModelsChange?: (deletedModel?: DeletedModelRef) => void;
onConfigure?: (id: string, meta: ModelSelectorChangeMeta) => void;
deleteDisabled?: boolean;
@@ -4229,6 +4355,111 @@ export function HubModelPicker({
const unpinRepo = usePinnedModelsStore((s) => s.unpinRepo);
const pinnedSet = useMemo(() => new Set(pinnedIds), [pinnedIds]);
+ // Connected pins keep their own list: an `external::` id carries the "::" the On Device store
+ // reads as its repo/quant separator, so a pin filed there returns as a phantom quant row.
+ const pinnedConnectedIds = usePinnedConnectedModelsStore((s) => s.pinned);
+ const togglePinnedConnected = usePinnedConnectedModelsStore(
+ (s) => s.togglePinnedConnected,
+ );
+ const movePinnedConnected = usePinnedConnectedModelsStore(
+ (s) => s.movePinnedConnected,
+ );
+ const beginPinnedConnectedDrag = usePinnedConnectedModelsStore(
+ (s) => s.beginPinnedConnectedDrag,
+ );
+ const endPinnedConnectedDrag = usePinnedConnectedModelsStore(
+ (s) => s.endPinnedConnectedDrag,
+ );
+ // The id under the cursor mid-drag. A ref, not state: dragenter fires on every row crossed.
+ const draggingPinnedConnectedRef = useRef<string | null>(null);
+ const [draggingPinnedConnectedId, setDraggingPinnedConnectedId] = useState<
+ string | null
+ >(null);
+ const pinnedConnectedSet = useMemo(
+ () => new Set(pinnedConnectedIds),
+ [pinnedConnectedIds],
+ );
+ // Connections behind those models, read for the base URL a capability check needs: a Gemini
+ // connection pointed at an OpenAI-compatible proxy returns no inline images.
+ const externalProviders = useExternalProvidersStore((s) => s.providers);
+ const externalBaseUrlById = useMemo(
+ () =>
+ new Map(
+ externalProviders.map((provider) => [provider.id, provider.baseUrl]),
+ ),
+ [externalProviders],
+ );
+ // A connection's own output cap, which lowers the model's documented one. The bounds the
+ // per-model editor offers have to be the ones every request is clamped to.
+ const externalMaxOutputById = useMemo(
+ () =>
+ new Map(
+ externalProviders.map((provider) => [
+ provider.id,
+ provider.maxOutputTokens ?? null,
+ ]),
+ ),
+ [externalProviders],
+ );
+ // A self-hosted OpenAI-compatible endpoint publishes no reasoning signal, so a vLLM connection
+ // carries the answer itself. It is the only place that answer lives.
+ const externalReasoningFlagById = useMemo(
+ () =>
+ new Map(
+ externalProviders.map((provider) => [
+ provider.id,
+ provider.isReasoningModel === true,
+ ]),
+ ),
+ [externalProviders],
+ );
+ // A provider catalogue arrives after first paint and decides most of the marks, so re-read it.
+ const catalogVersion = useSyncExternalStore(
+ subscribeModelCatalog,
+ modelCatalogVersion,
+ );
+ const [connectedSort, setConnectedSort] =
+ useState<ConnectedSortKey>("provider");
+ const [connectedModality, setConnectedModality] =
+ useState<ConnectedModalityFilter>("all");
+ // Provider ids the user folded away, and whether the Pinned group is folded.
+ const [collapsedConnectedGroups, setCollapsedConnectedGroups] = useState<
+ ReadonlySet<string>
+ >(() => new Set());
+ const [pinnedConnectedCollapsed, setPinnedConnectedCollapsed] =
+ useState(false);
+ const toggleConnectedGroup = useCallback((providerId: string) => {
+ setCollapsedConnectedGroups((prev) => {
+ const next = new Set(prev);
+ if (!next.delete(providerId)) next.add(providerId);
+ return next;
+ });
+ }, []);
+ // The row whose Model info is open.
+ const [infoModel, setInfoModel] = useState<{
+ model: ExternalModelOption;
+ providerModelId: string;
+ baseUrl: string | null;
+ isReasoningProvider: boolean;
+ } | null>(null);
+ const [settingsModel, setSettingsModel] = useState<{
+ model: ExternalModelOption;
+ providerModelId: string;
+ baseUrl: string | null;
+ isReasoningProvider: boolean;
+ connectionMaxOutputTokens: number | null;
+ } | null>(null);
+
+ // The shared helper, not navigator.clipboard: that is undefined in the desktop shell and over
+ // plain HTTP on a LAN address, where reading .writeText off it throws before any catch runs.
+ const copyConnectedModelId = useCallback(async (providerModelId: string) => {
+ if (await copyToClipboard(providerModelId)) {
+ toast.success(`Copied ${providerModelId}`);
+ } else {
+ toast.error("Could not copy the model ID");
+ }
+ }, []);
+
// Candidate pins whose repo still exists in the cache; per-quant validation below is needed
// because deleting one variant can leave a sibling cached.
const pinnedQuantCandidates = useMemo(() => {
@@ -4545,9 +4776,116 @@ export function HubModelPicker({
[hfIds, communitySearchIds],
);
+ // Query- and modality-matching connected models, pinned ones kept separate. A pin moves a row
+ // into the Pinned group rather than copying it, as the On Device sections do.
+ const connectedMatches = useMemo(() => {
+ const needle = normalizeForSearch(debouncedQuery.trim());
+ return externalModels.filter((model) => {
+ if (
+ needle &&
+ !normalizeForSearch(
+ `${model.name} ${model.providerName} ${model.id}`,
+ ).includes(needle)
+ ) {
+ return false;
+ }
+ if (connectedModality === "all") return true;
+ const marks = connectedModelMarks({
+ providerType: model.providerType,
+ modelId: parseExternalModelId(model.id)?.modelId ?? model.name,
+ baseUrl: externalBaseUrlById.get(model.providerId) ?? null,
+ });
+ return connectedModality === "vision"
+ ? marks.vision
+ : marks.capabilities[connectedModality];
+ });
+ // The marks read module state a provider sync writes after first paint, so the version is
+ // the only dep that can report the memo stale. exhaustive-deps cannot see through
+ // connectedModelMarks to that read, so it calls the dep unnecessary.
+ // eslint-disable-next-line react-hooks/exhaustive-deps
+ }, [
+ externalModels,
+ debouncedQuery,
+ connectedModality,
+ externalBaseUrlById,
+ catalogVersion,
+ ]);
+
+ // Pin order, always: the sort below orders the catalogue, and a hand-arranged list is not it.
+ const pinnedConnectedRows = useMemo(() => {
+ const rank = makePinRank(pinnedConnectedIds);
+ return connectedMatches
+ .filter((model) => pinnedConnectedSet.has(model.id))
+ .sort((a, b) => rank(a.id) - rank(b.id));
+ }, [connectedMatches, pinnedConnectedIds, pinnedConnectedSet]);
+
+ const connectedGroups = useMemo(() => {
+ const byProvider = new Map<
+ string,
+ {
+ providerId: string;
+ providerName: string;
+ providerType: string;
+ models: ExternalModelOption[];
+ }
+ >();
+ for (const model of connectedMatches) {
+ if (pinnedConnectedSet.has(model.id)) continue;
+ const prev = byProvider.get(model.providerId);
+ if (prev) {
+ prev.models.push(model);
+ } else {
+ byProvider.set(model.providerId, {
+ providerId: model.providerId,
+ providerName: model.providerName,
+ providerType: model.providerType,
+ models: [model],
+ });
+ }
+ }
+ const groups = [...byProvider.values()]
+ .map((group) => ({
+ ...group,
+ models: group.models.sort((a, b) => a.name.localeCompare(b.name)),
+ }))
+ .sort((a, b) => a.providerName.localeCompare(b.providerName));
+ if (connectedSort !== "name") return groups;
+ // Sorting by name across connections is a different list, not a reordered one: the headings
+ // would each hold a row or two and stop meaning anything. So one unlabelled group.
+ return [
+ {
+ providerId: "__all__",
+ providerName: "",
+ providerType: "",
+ models: groups
+ .flatMap((group) => group.models)
+ .sort((a, b) => a.name.localeCompare(b.name)),
+ },
+ ];
+ }, [connectedMatches, pinnedConnectedSet, connectedSort]);
const hubOptionKeys = useMemo(() => {
const keys: string[] = [];
+ // The tab lists nothing else, so these are the whole roving order, in drawn order.
+ if (section === "connected") {
+ if (!pinnedConnectedCollapsed) {
+ keys.push(
+ ...pinnedConnectedRows.map((model) =>
+ makeModelOptionKey("connected", model.id),
+ ),
+ );
+ }
+ for (const group of connectedGroups) {
+ if (collapsedConnectedGroups.has(group.providerId)) continue;
+ keys.push(
+ ...group.models.map((model) =>
+ makeModelOptionKey("connected", model.id),
+ ),
+ );
+ }
+ return keys;
+ }
+
// Pinned rows sit above the Unsloth heading on the On Device tab.
if (
section === "downloaded" &&
@@ -4671,6 +5009,10 @@ export function HubModelPicker({
chatOnly,
sortedCustomFolderModels,
customFoldersCollapsed,
+ connectedGroups,
+ pinnedConnectedRows,
+ pinnedConnectedCollapsed,
+ collapsedConnectedGroups,
pinnedRows,
pinnedCollapsed,
downloadedCollapsed,
@@ -5008,41 +5350,6 @@ export function HubModelPicker({
/>
);
- const connectedGroups = useMemo(() => {
- const needle = normalizeForSearch(debouncedQuery.trim());
- const byProvider = new Map<
- string,
- {
- providerId: string;
- providerName: string;
- providerType: string;
- models: ExternalModelOption[];
- }
- >();
- for (const model of externalModels) {
- const text = normalizeForSearch(
- `${model.name} ${model.providerName} ${model.id}`,
- );
- if (needle && !text.includes(needle)) continue;
- const prev = byProvider.get(model.providerId);
- if (prev) {
- prev.models.push(model);
- } else {
- byProvider.set(model.providerId, {
- providerId: model.providerId,
- providerName: model.providerName,
- providerType: model.providerType,
- models: [model],
- });
- }
- }
- return [...byProvider.values()]
- .map((group) => ({
- ...group,
- models: group.models.sort((a, b) => a.name.localeCompare(b.name)),
- }))
- .sort((a, b) => a.providerName.localeCompare(b.providerName));
- }, [externalModels, debouncedQuery]);
const showConnected = section === "connected";
// The Connected layout uses a wider box, so it drops the search inset to keep Search Hub on the last dropdown's edge.
const hasConnected = externalModels.length > 0;
@@ -5067,6 +5374,187 @@ export function HubModelPicker({
selected && "bg-[#ececec] dark:bg-[var(--sidebar-accent)]",
);
+ // One connected model, through ModelRow like every On Device row, so the badges and the hover
+ // gutter are the same components rather than a second set that drifts.
+ const renderConnectedModelRow = (
+ model: ExternalModelOption,
+ // Only pinned rows drag: the groups below are sorted, so a drop there could not be honoured.
+ draggable = false,
+ // No heading above this row to name its connection: the pinned group, and the flat list the
+ // name sort produces. Two connections can serve one model id, so the row has to say.
+ headless = false,
+ ) => {
+ const optionKey = makeModelOptionKey("connected", model.id);
+ const isSelected = value === model.id;
+ const isPinned = pinnedConnectedSet.has(model.id);
+ // What the provider calls the model: `model.name` is a display label OpenRouter rewrites and
+ // `model.id` is the picker's `external::` address, and no capability lookup keys on either.
+ const providerModelId =
+ parseExternalModelId(model.id)?.modelId ?? model.name;
+ const baseUrl = externalBaseUrlById.get(model.providerId) ?? null;
+ const marks = connectedModelMarks({
+ providerType: model.providerType,
+ modelId: providerModelId,
+ baseUrl,
+ });
+ return (
+ <div
+ key={model.id}
+ // ml-4 plus the pl-3.5 below is 30px, where a heading's label starts: px-2.5 + a
+ // size-3.5 icon + gap-1.5. The split between them is the pill's own inset, so widening
+ // it takes the pill's left edge leftward rather than moving the name off that label.
+ // The reserved leading slot used to put 23.5px of empty pill in front of every name.
+ className={cn(downloadedRowShellClassName(isSelected), "ml-4")}
+ style={
+ draggingPinnedConnectedId === model.id ? { opacity: 0.4 } : undefined
+ }
+ draggable={draggable}
+ onDragStart={
+ draggable
+ ? (event) => {
+ event.dataTransfer.effectAllowed = "move";
+ // Firefox will not start a drag without data.
+ event.dataTransfer.setData("text/plain", model.id);
+ draggingPinnedConnectedRef.current = model.id;
+ setDraggingPinnedConnectedId(model.id);
+ // Reordering is live on dragenter; this is what a cancelled drag rolls back to.
+ beginPinnedConnectedDrag();
+ }
+ : undefined
+ }
+ onDragEnd={
+ draggable
+ ? () => {
+ draggingPinnedConnectedRef.current = null;
+ setDraggingPinnedConnectedId(null);
+ // Escape, or a release off-row, reaches dragend without a drop. After a drop the
+ // session is already committed and cleared, so this is a no-op.
+ endPinnedConnectedDrag(false);
+ }
+ : undefined
+ }
+ onDragOver={
+ draggable
+ ? (event) => {
+ if (draggingPinnedConnectedRef.current) event.preventDefault();
+ }
+ : undefined
+ }
+ onDragEnter={
+ draggable
+ ? () => {
+ const dragId = draggingPinnedConnectedRef.current;
+ if (dragId && dragId !== model.id) {
+ movePinnedConnected(dragId, model.id);
+ }
+ }
+ : undefined
+ }
+ onDrop={
+ draggable
+ ? (event) => {
+ event.preventDefault();
+ draggingPinnedConnectedRef.current = null;
+ setDraggingPinnedConnectedId(null);
+ endPinnedConnectedDrag(true);
+ }
+ : undefined
+ }
+ >
+ <div className="min-w-0 flex-1">
+ <ModelRow
+ label={model.name}
+ // The provider's own id, the way a local row hovers its path: a label can be
+ // rewritten, and one connection can offer two models that shorten to the same words.
+ // With no heading above, the connection's name goes here too.
+ tooltipText={
+ headless ? (
+ <>
+ {providerModelId}
+ <span className="block text-ui-10 mt-1">
+ {model.providerName}
+ </span>
+ </>
+ ) : (
+ providerModelId
+ )
+ }
+ capabilities={marks.capabilities}
+ showVision={marks.vision}
+ selected={isSelected}
+ optionProps={hubModelList.getOptionProps(optionKey, isSelected)}
+ onClick={() =>
+ onSelect(model.id, { source: "external", isLora: false })
+ }
+ vramStatus={null}
+ // The name's own inset, since nothing precedes it in the row now: the leading slot a
+ // local row gives its format dot is gone with the logo that briefly filled it.
+ className={cn(downloadedRowButtonClassName, "pl-3.5")}
+ />
+ </div>
+ <span className={ROW_ACTIONS_CLASS}>
+ <ModelLoadSettingsAction
+ ariaLabel={`Settings for ${model.name}`}
+ tooltip="Model settings"
+ onConfigure={() =>
+ setSettingsModel({
+ model,
+ providerModelId,
+ baseUrl,
+ isReasoningProvider:
+ externalReasoningFlagById.get(model.providerId) === true,
+ connectionMaxOutputTokens:
+ externalMaxOutputById.get(model.providerId) ?? null,
+ })
+ }
+ />
+ <ModelRowMenu
+ ariaLabel={`More options for ${model.name}`}
+ pin={{
+ pinned: isPinned,
+ pinLabel: "Pin to top",
+ unpinLabel: "Unpin",
+ onToggle: () => togglePinnedConnected(model.id),
+ }}
+ items={[
+ {
+ key: "info",
+ label: "Model info",
+ icon: (
+ <HugeiconsIcon
+ icon={InformationCircleIcon}
+ strokeWidth={1.75}
+ className="size-icon"
+ />
+ ),
+ onSelect: () =>
+ setInfoModel({
+ model,
+ providerModelId,
+ baseUrl,
+ isReasoningProvider:
+ externalReasoningFlagById.get(model.providerId) === true,
+ }),
+ },
+ {
+ key: "copy",
+ label: "Copy model ID",
+ icon: (
+ <HugeiconsIcon
+ icon={Copy01Icon}
+ strokeWidth={1.75}
+ className="size-icon"
+ />
+ ),
+ onSelect: () => void copyConnectedModelId(providerModelId),
+ },
+ ]}
+ />
+ </span>
+ </div>
+ );
+ };
+
// A pinned quant: repo name with the quant as a grey chip, loaded in one click.
const renderPinnedQuantRow = (entry: { repoId: string; quant: string }) => {
const optionKey = makeModelOptionKey(
@@ -5661,7 +6149,33 @@ export function HubModelPicker({
)}
>
{sectionToggle}
- {showConnected ? null : (
+ {showConnected ? (
+ <div className="flex max-w-full min-w-0 flex-wrap items-center gap-2">
+ <HubOptionMenu
+ value={connectedModality}
+ options={CONNECTED_MODALITY_OPTIONS}
+ onValueChange={setConnectedModality}
+ ariaLabel="Filter by modality"
+ align="end"
+ className={sortTriggerClassName}
+ contentClassName={sortMenuContentClassName}
+ />
+ <HubOptionMenu
+ value={connectedSort}
+ options={CONNECTED_SORT_OPTIONS}
+ onValueChange={setConnectedSort}
+ ariaLabel="Sort connected models"
+ align="end"
+ className={sortTriggerClassName}
+ contentClassName={sortMenuContentClassName}
+ triggerContent={sortTriggerContent(
+ CONNECTED_SORT_OPTIONS.find(
+ (option) => option.value === connectedSort,
+ )?.label ?? connectedSort,
+ )}
+ />
+ </div>
+ ) : (
<div className="flex max-w-full min-w-0 flex-wrap items-center gap-2">
<HubOptionMenu
value={formatFilter}
@@ -5700,47 +6214,73 @@ export function HubModelPicker({
)}
>
{showConnected ? (
- connectedGroups.length === 0 ? (
+ connectedMatches.length === 0 ? (
<div className="px-2.5 py-2 text-xs leading-relaxed text-muted-foreground">
{externalModels.length === 0
? "No models from your connections. Set up in Settings then Connections."
: "No models match your search."}
</div>
) : (
- connectedGroups.map((group) => (
- <div key={group.providerId}>
- {/* Wider than the On Device section labels: nothing divides these groups but the gap. */}
- <div className="flex items-center gap-2 px-2.5 pb-1 pt-5 text-ui-10 font-semibold uppercase tracking-wider text-muted-foreground">
- <ApiProviderLogo
- providerType={group.providerType}
- className="size-3.5"
- title={group.providerName}
+ <>
+ {/* Above the first provider, on the same heading with a pin for the logo. */}
+ {pinnedConnectedRows.length > 0 ? (
+ <div>
+ <ConnectedGroupHeading
+ icon={
+ <HugeiconsIcon icon={PinIcon} className="size-3.5" />
+ }
+ label="Pinned"
+ collapsed={pinnedConnectedCollapsed}
+ onToggle={() =>
+ setPinnedConnectedCollapsed((value) => !value)
+ }
/>
- <span className="min-w-0 truncate">
- {group.providerName}
- </span>
+ {pinnedConnectedCollapsed
+ ? null
+ : pinnedConnectedRows.map((model) =>
+ renderConnectedModelRow(model, true, true),
+ )}
</div>
- {group.models.map((model) => (
- <button
- key={model.id}
- type="button"
- onClick={() =>
- onSelect(model.id, {
- source: "external",
- isLora: false,
- })
- }
- className={cn(
- "flex w-full items-center rounded-md px-2.5 py-1.5 text-left text-sm transition-colors hover:bg-[#ececec] dark:hover:bg-[var(--sidebar-accent)]",
- value === model.id &&
- "bg-[#ececec] dark:bg-[var(--sidebar-accent)]",
- )}
- >
- <span className="min-w-0 truncate">{model.name}</span>
- </button>
- ))}
- </div>
- ))
+ ) : null}
+ {connectedGroups.map((group) => {
+ // The name sort collapses every connection into one unlabelled group, which
+ // has no heading to fold.
+ const headed = group.providerName.length > 0;
+ const collapsed =
+ headed && collapsedConnectedGroups.has(group.providerId);
+ return (
+ <div key={group.providerId}>
+ {headed ? (
+ <ConnectedGroupHeading
+ icon={
+ <ApiProviderLogo
+ providerType={group.providerType}
+ className="size-3.5"
+ title={group.providerName}
+ />
+ }
+ label={group.providerName}
+ collapsed={collapsed}
+ onToggle={() =>
+ toggleConnectedGroup(group.providerId)
+ }
+ onConfigure={
+ onConfigureConnection
+ ? () => onConfigureConnection(group.providerId)
+ : undefined
+ }
+ configureLabel={`${group.providerName} connection settings`}
+ />
+ ) : null}
+ {collapsed
+ ? null
+ : group.models.map((model) =>
+ renderConnectedModelRow(model, false, !headed),
+ )}
+ </div>
+ );
+ })}
+ </>
)
) : (
<>
@@ -6980,6 +7520,35 @@ export function HubModelPicker({
onKeepTransport={resumeUpdateConflict}
onSwitchTransport={restartUpdateConflict}
/>
+ {settingsModel ? (
+ <ConnectedModelSettingsDialog
+ open={true}
+ onOpenChange={(next) => {
+ if (!next) setSettingsModel(null);
+ }}
+ checkpointId={settingsModel.model.id}
+ displayName={settingsModel.model.name}
+ modelId={settingsModel.providerModelId}
+ providerType={settingsModel.model.providerType}
+ baseUrl={settingsModel.baseUrl}
+ isReasoningProvider={settingsModel.isReasoningProvider}
+ connectionMaxOutputTokens={settingsModel.connectionMaxOutputTokens}
+ />
+ ) : null}
+ {infoModel ? (
+ <ConnectedModelInfoDialog
+ open={true}
+ onOpenChange={(next) => {
+ if (!next) setInfoModel(null);
+ }}
+ modelId={infoModel.providerModelId}
+ displayName={infoModel.model.name}
+ providerName={infoModel.model.providerName}
+ providerType={infoModel.model.providerType}
+ baseUrl={infoModel.baseUrl}
+ isReasoningProvider={infoModel.isReasoningProvider}
+ />
+ ) : null}
</CapabilityScope.Provider>
);
}
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/pinned-connected-models.ts b/studio/frontend/src/features/model-picker/components/model-selector/pinned-connected-models.ts
new file mode 100644
index 00000000000..a83b753db09
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/components/model-selector/pinned-connected-models.ts
@@ -0,0 +1,154 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Pinned `external::<connectionId>::<modelId>` ids for the Connected list, in localStorage.
+// NOT the On Device store in pinned-models.ts: an external id contains "::", which that store
+// reads as its repoId/quant separator and surfaces as a phantom pinned quant on the On Device tab.
+
+import { create } from "zustand";
+
+const KEY = "unsloth_pinned_connected_models";
+
+function readPinned(): string[] {
+ try {
+ const raw = JSON.parse(localStorage.getItem(KEY) ?? "[]");
+ return Array.isArray(raw)
+ ? raw.filter((v): v is string => typeof v === "string")
+ : [];
+ } catch {
+ return [];
+ }
+}
+
+/** The stored list, or null for "nothing to read". Distinct from []: a toggle falling back to []
+ * would drop this window's own pins where every write has failed. */
+function storedPinned(): string[] | null {
+ try {
+ const raw = localStorage.getItem(KEY);
+ if (raw === null) return null;
+ const parsed = JSON.parse(raw);
+ return Array.isArray(parsed)
+ ? parsed.filter((v): v is string => typeof v === "string")
+ : null;
+ } catch {
+ return null;
+ }
+}
+
+function writePinned(pinned: string[]): void {
+ try {
+ localStorage.setItem(KEY, JSON.stringify(pinned));
+ } catch {
+ // Ignore unavailable storage; pins stay session-only.
+ }
+}
+
+// movePinnedConnected runs on every dragenter: writing each would make a cancelled drag permanent.
+// Snapshot, move in memory, commit on drop. Mirrors pinned-models.ts.
+let dragSnapshot: string[] | null = null;
+
+// Keep remote updates for cancellation without replacing the drag's live order.
+let dragExternalOrder: string[] | null = null;
+
+function sameOrder(a: readonly string[], b: readonly string[]): boolean {
+ return a.length === b.length && a.every((key, index) => key === b[index]);
+}
+
+/** A dragged order over the stored membership: this window owns row ORDER, the record owns which
+ * ids are pinned, since another window's storage event may still be in flight at the drop.
+ * Its additions arrive at the front; null storage leaves the order alone. */
+function rebaseOnStored(order: readonly string[]): string[] {
+ const stored = storedPinned();
+ if (stored === null) return [...order];
+ const kept = order.filter((id) => stored.includes(id));
+ const added = stored.filter((id) => !order.includes(id));
+ return [...added, ...kept];
+}
+
+interface PinnedConnectedModelsState {
+ pinned: string[];
+ /** Pin or unpin one external model id. */
+ togglePinnedConnected: (modelId: string) => void;
+ /**
+ * Move `fromId` into `toId`'s slot. Both must already be pinned; anything else is a no-op.
+ * Outside a drag session the new order persists immediately, inside one it is held until
+ * `endPinnedConnectedDrag`.
+ */
+ movePinnedConnected: (fromId: string, toId: string) => void;
+ /** Snapshot the current order so a cancelled drag can be undone. */
+ beginPinnedConnectedDrag: () => void;
+ /** End a drag session. `commit` persists the reordered list; otherwise the snapshot is
+ * restored. Idempotent, because drop is followed by dragend and only the first may decide. */
+ endPinnedConnectedDrag: (commit: boolean) => void;
+}
+
+export const usePinnedConnectedModelsStore = create<PinnedConnectedModelsState>(
+ (set) => ({
+ pinned: readPinned(),
+ togglePinnedConnected: (modelId) =>
+ set((state) => {
+ // Applied to the stored list, not to this window's copy of it, since a write replaces the
+ // whole list: another window's pin can be newer than the storage event this one has
+ // processed, and rewriting our own array would drop it for good.
+ const base = storedPinned() ?? state.pinned;
+ // Newest pin first, as On Device does, so "Pin to top" literally lands on top of the
+ // pinned group rather than under earlier pins.
+ const next = base.includes(modelId)
+ ? base.filter((id) => id !== modelId)
+ : [modelId, ...base];
+ writePinned(next);
+ return { pinned: next };
+ }),
+ movePinnedConnected: (fromId, toId) =>
+ set((state) => {
+ const from = state.pinned.indexOf(fromId);
+ const to = state.pinned.indexOf(toId);
+ if (from < 0 || to < 0 || from === to) return state;
+ const next = [...state.pinned];
+ next.splice(to, 0, ...next.splice(from, 1));
+ if (dragSnapshot === null) writePinned(next);
+ return { pinned: next };
+ }),
+ beginPinnedConnectedDrag: () =>
+ set((state) => {
+ dragSnapshot = [...state.pinned];
+ dragExternalOrder = null;
+ return state;
+ }),
+ endPinnedConnectedDrag: (commit) =>
+ set((state) => {
+ const snapshot = dragSnapshot;
+ const external = dragExternalOrder;
+ dragSnapshot = null;
+ dragExternalOrder = null;
+ if (snapshot === null) return state;
+ // Read storage too, since its event may still be pending.
+ const base = storedPinned() ?? external ?? snapshot;
+ if (commit && !sameOrder(snapshot, state.pinned)) {
+ if (sameOrder(base, state.pinned)) return state;
+ // The write replaces the whole list and nothing echoes it back to this window, so a
+ // plain write of the dragged order would erase a pin another window added mid-drag
+ // with no event left to restore it.
+ const next = rebaseOnStored(state.pinned);
+ writePinned(next);
+ return sameOrder(next, state.pinned) ? state : { pinned: next };
+ }
+ // Cancellation and unchanged drags use the latest persisted order.
+ if (sameOrder(base, state.pinned)) return state;
+ return { pinned: base };
+ }),
+ }),
+);
+
+if (typeof window !== "undefined") {
+ window.addEventListener("storage", (event) => {
+ if (event.key === KEY || event.key === null) {
+ const next = readPinned();
+ if (dragSnapshot !== null) {
+ dragExternalOrder = next;
+ return;
+ }
+ usePinnedConnectedModelsStore.setState({ pinned: next });
+ }
+ });
+}
diff --git a/studio/frontend/src/features/model-picker/index.ts b/studio/frontend/src/features/model-picker/index.ts
index 92b2e9de242..5a4e2306c85 100644
--- a/studio/frontend/src/features/model-picker/index.ts
+++ b/studio/frontend/src/features/model-picker/index.ts
@@ -10,6 +10,10 @@ export {
pinKey,
usePinnedModelsStore,
} from "./components/model-selector/pinned-models";
+export {
+ pinnedReasoningEffort,
+ useModelReasoningEffortStore,
+} from "./components/model-selector/model-reasoning-effort";
export {
hfModelFitsDevice,
loadScopedGpu,
diff --git a/studio/frontend/src/features/settings/stores/settings-dialog-store.ts b/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
index 539eaeec83f..46bcdf51b4b 100644
--- a/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
+++ b/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
@@ -54,10 +54,16 @@ interface SettingsDialogState {
// toast). DataTab uses it as its initial subpage, then clears it. See requestsFor
// for how long it lives unconsumed.
archivedRequested: ArchivedShelf | null;
+ // Set when something asks for one connection's settings (the picker's Connected group gear).
+ // ConnectionsTab hands it to the form, then clears it. Same lifetime as archivedRequested.
+ connectionRequested: string | null;
openDialog: (tab?: SettingsTab, options?: OpenDialogOptions) => void;
openArchivedChats: () => void;
openArchivedMedia: (shelf: Exclude<ArchivedShelf, "chats">) => void;
+ /** Open Connections with `providerId`'s edit form already up. */
+ openConnectionSettings: (providerId: string) => void;
consumeArchivedChatsRequest: () => void;
+ consumeConnectionRequest: () => void;
consumeScrollTarget: (target: SettingsScrollTarget) => void;
closeDialog: () => void;
setActiveTab: (tab: SettingsTab) => void;
@@ -130,6 +136,8 @@ function requestsFor(state: SettingsDialogState, tab: SettingsTab) {
? state.scrollTarget
: null,
archivedRequested: tab === "data" ? state.archivedRequested : null,
+ connectionRequested:
+ tab === "connections" ? state.connectionRequested : null,
};
}
@@ -140,6 +148,7 @@ export const useSettingsDialogStore = create<SettingsDialogState>((set) => ({
opener: null,
openerFallback: null,
archivedRequested: null,
+ connectionRequested: null,
openDialog: (tab, options) =>
set((state) => {
const next = tab ?? state.activeTab;
@@ -150,6 +159,7 @@ export const useSettingsDialogStore = create<SettingsDialogState>((set) => ({
// A caller that names a target replaces whatever was still pending.
scrollTarget: options?.scrollTarget ?? pending.scrollTarget,
archivedRequested: pending.archivedRequested,
+ connectionRequested: pending.connectionRequested,
...focusForOpen(state, options?.focusFallback),
};
}),
@@ -159,6 +169,7 @@ export const useSettingsDialogStore = create<SettingsDialogState>((set) => ({
activeTab: "data",
scrollTarget: null,
archivedRequested: "chats",
+ connectionRequested: null,
...focusForOpen(state),
})),
openArchivedMedia: (shelf) =>
@@ -167,9 +178,20 @@ export const useSettingsDialogStore = create<SettingsDialogState>((set) => ({
activeTab: "data",
scrollTarget: null,
archivedRequested: shelf,
+ connectionRequested: null,
+ ...focusForOpen(state),
+ })),
+ openConnectionSettings: (providerId) =>
+ set((state) => ({
+ open: true,
+ activeTab: "connections",
+ scrollTarget: null,
+ archivedRequested: null,
+ connectionRequested: providerId,
...focusForOpen(state),
})),
consumeArchivedChatsRequest: () => set({ archivedRequested: null }),
+ consumeConnectionRequest: () => set({ connectionRequested: null }),
consumeScrollTarget: (target) =>
set((state) => ({
scrollTarget: state.scrollTarget === target ? null : state.scrollTarget,
@@ -178,7 +200,12 @@ export const useSettingsDialogStore = create<SettingsDialogState>((set) => ({
// pass after `open: false` lands, so the opener must still be readable
// from the store at that point. The next openDialog() overwrites it.
closeDialog: () =>
- set({ open: false, scrollTarget: null, archivedRequested: null }),
+ set({
+ open: false,
+ scrollTarget: null,
+ archivedRequested: null,
+ connectionRequested: null,
+ }),
setActiveTab: (tab) => {
try {
window.localStorage.setItem(ACTIVE_TAB_KEY, tab);
diff --git a/studio/frontend/src/features/settings/tabs/connections-tab.tsx b/studio/frontend/src/features/settings/tabs/connections-tab.tsx
index a8b8b598664..a056a0e109d 100644
--- a/studio/frontend/src/features/settings/tabs/connections-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/connections-tab.tsx
@@ -3,15 +3,25 @@
import { ChatProvidersSettings } from "@/features/chat/chat-providers-dialog";
import { useExternalProvidersStore } from "@/features/chat/stores/external-providers-store";
+import { useSettingsDialogStore } from "../stores/settings-dialog-store";
export function ConnectionsTab() {
const providers = useExternalProvidersStore((s) => s.providers);
const setProviders = useExternalProvidersStore((s) => s.setProviders);
+ // Set when the picker's Connected group gear asked for one connection by name.
+ const connectionRequested = useSettingsDialogStore(
+ (s) => s.connectionRequested,
+ );
+ const consumeConnectionRequest = useSettingsDialogStore(
+ (s) => s.consumeConnectionRequest,
+ );
return (
<ChatProvidersSettings
providers={providers}
onProvidersChange={setProviders}
+ openProviderId={connectionRequested}
+ onOpenProviderConsumed={consumeConnectionRequest}
/>
);
}
diff --git a/studio/frontend/tests/connected-model-row-actions.test.ts b/studio/frontend/tests/connected-model-row-actions.test.ts
new file mode 100644
index 00000000000..5764615f3dd
--- /dev/null
+++ b/studio/frontend/tests/connected-model-row-actions.test.ts
@@ -0,0 +1,813 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Connected rows used to be a name and nothing else, while every On Device row carried modality
+// badges, a pin and a gear. These pin the parts that closed that gap, and the one boundary it
+// must not be closed across: the two pin lists stay apart.
+
+import assert from "node:assert/strict";
+import test from "node:test";
+
+import { readSrc } from "./helpers/kit.ts";
+
+const pickers = readSrc(
+ "features/model-picker/components/model-selector/pickers.tsx",
+);
+const marks = readSrc(
+ "features/model-picker/components/model-selector/connected-model-meta.ts",
+);
+const connectedPins = readSrc(
+ "features/model-picker/components/model-selector/pinned-connected-models.ts",
+);
+const onDevicePins = readSrc(
+ "features/model-picker/components/model-selector/pinned-models.ts",
+);
+const selector = readSrc("features/model-picker/components/model-selector.tsx");
+const settingsStore = readSrc(
+ "features/settings/stores/settings-dialog-store.ts",
+);
+const connectionsTab = readSrc("features/settings/tabs/connections-tab.tsx");
+const providersDialog = readSrc("features/chat/chat-providers-dialog.tsx");
+
+test("a connected row draws its badges through ModelRow", () => {
+ // The same component the On Device rows use, so the glyph set cannot drift into a second one.
+ assert.match(
+ pickers,
+ /const renderConnectedModelRow = \(\s*model: ExternalModelOption,/,
+ );
+ assert.match(
+ pickers,
+ /capabilities=\{marks\.capabilities\}\s*\n\s*showVision=\{marks\.vision\}/,
+ );
+ // No format-dot slot: nothing precedes a connected name, and an empty one left the hover pill
+ // starting 23.5px in front of it. The row's own pl-3.5 is the name's inset instead.
+ assert.doesNotMatch(pickers, /reserveLeadingSlot/);
+ // The connection is named by the group heading, so it is not repeated on each of its rows.
+ assert.doesNotMatch(pickers, /leadingBadge=\{\s*<ApiProviderLogo/);
+ // Both groups render through it. Only the pinned one drags: the groups below are sorted, so a
+ // drop there could not be honoured.
+ assert.match(
+ pickers,
+ /pinnedConnectedRows\.map\(\(model\) =>\s*renderConnectedModelRow\(model, true, true\)/,
+ );
+ assert.match(
+ pickers,
+ /group\.models\.map\(\(model\) =>\s*renderConnectedModelRow\(model, false, !headed\)/,
+ );
+});
+
+test("capabilities are keyed by the provider's own model id", () => {
+ // `model.id` is the external:: address and `model.name` a label OpenRouter rewrites, so a
+ // lookup on either misses every catalogue entry.
+ assert.match(
+ pickers,
+ /parseExternalModelId\(model\.id\)\?\.modelId \?\? model\.name/,
+ );
+ assert.match(
+ pickers,
+ /connectedModelMarks\(\{\s*providerType: model\.providerType,\s*modelId: providerModelId,\s*baseUrl,/,
+ );
+ // A catalogue lands after first paint, so the marks have to be re-read when it does.
+ assert.match(
+ pickers,
+ /useSyncExternalStore\(\s*subscribeModelCatalog,\s*modelCatalogVersion,?\s*\)/,
+ );
+});
+
+test("modality comes from the resolvers the app already has", () => {
+ assert.match(
+ marks,
+ /providerModelSupportsVision\(providerType, modelId\) === true/,
+ );
+ assert.match(marks, /providerSupportsBuiltinImageGeneration\(/);
+ // Video generation is never claimed on a connected row: no provider publishes a capability for
+ // it and nothing we connect to serves it through the chat route, so the name would be the only
+ // evidence, and the glyph would promise a row something selecting it cannot do.
+ assert.match(marks, /videoGen: false,/);
+ // Nor is any other mark guessed from the name. Image generation is enabled in the adapter
+ // through providerSupportsBuiltinImageGeneration alone, and the composer's Images pill with it,
+ // so a name that reads like a diffusion model but resolves to false there can never be asked
+ // for one.
+ assert.doesNotMatch(marks, /byName/);
+ assert.doesNotMatch(marks, /detectCapabilities/);
+ assert.match(
+ marks,
+ /imageGen: providerSupportsBuiltinImageGeneration\(\s*providerType,\s*modelId,\s*baseUrl,\s*\),/,
+ );
+ // Audio is withheld even where a catalogue publishes it as an input modality: the attachment
+ // adapter resolves the active model out of `models`, which carries loaded local models only, so
+ // an external selection has no row there and `add` rejects every audio file it is given.
+ assert.match(marks, /audio: false,/);
+ assert.doesNotMatch(marks, /includes\("audio"\)/);
+ const audioAdapter = readSrc("features/chat/audio-attachment-adapter.ts");
+ assert.match(
+ audioAdapter,
+ /const activeModel = state\.models\.find\(\(m\) => m\.id === checkpoint\);/,
+ );
+ assert.match(audioAdapter, /\} else if \(!activeModel\?\.hasAudioInput\) \{/);
+ // So the filter drops its Audio option rather than offering one that matches nothing.
+ assert.match(
+ pickers,
+ /type ConnectedModalityFilter = "all" \| "vision" \| "imageGen";/,
+ );
+ // Unknown is not a promise: only an explicit true draws the eye.
+ assert.doesNotMatch(
+ marks,
+ /vision: providerModelSupportsVision\([^)]*\) \?\?/,
+ );
+});
+
+test("connected pins are a separate list from the On Device ones", () => {
+ assert.match(connectedPins, /"unsloth_pinned_connected_models"/);
+ assert.match(onDevicePins, /"unsloth_pinned_models"/);
+ // An external id carries "::", which this reads as the repo/quant separator, so a pin filed in
+ // the On Device store comes back out as a phantom pinned quant on that tab.
+ assert.match(onDevicePins, /const sep = key\.indexOf\("::"\)/);
+ assert.doesNotMatch(connectedPins, /"unsloth_pinned_models"/);
+ // So a connected row reaches for its own toggle and never the On Device one.
+ assert.doesNotMatch(pickers, /togglePinned\(model\.id\)/);
+ assert.doesNotMatch(pickers, /pinKey\(model\.id\)/);
+ assert.match(pickers, /onToggle: \(\) => togglePinnedConnected\(model\.id\)/);
+ // A toggle applies its one change to the STORED list. The write replaces the whole list, so two
+ // windows pinning different models before the storage event lands would otherwise keep only the
+ // second pin. Null, not [], when there is nothing to read: a toggle on an install whose writes
+ // all fail has to keep the session's own pins rather than drop them.
+ assert.match(
+ connectedPins,
+ /const base = storedPinned\(\) \?\? state\.pinned;/,
+ );
+ assert.match(
+ connectedPins,
+ /const raw = localStorage\.getItem\(KEY\);\s*if \(raw === null\) return null;/,
+ );
+ // A drag commit rebases too, for the same reason and one more: nothing echoes this window's own
+ // write back to it, so a pin another window added mid-drag would be erased with no event left
+ // to restore it. The dragged order is this window's; which ids are pinned is the record's.
+ assert.match(
+ connectedPins,
+ /const next = rebaseOnStored\(state\.pinned\);\s*writePinned\(next\);/,
+ );
+ assert.match(
+ connectedPins,
+ /const kept = order\.filter\(\(id\) => stored\.includes\(id\)\);\s*const added = stored\.filter\(\(id\) => !order\.includes\(id\)\);\s*return \[\.\.\.added, \.\.\.kept\];/,
+ );
+});
+
+test("a pin moves the row out of its provider group", () => {
+ // Copying it instead would list the model twice under one search.
+ assert.match(
+ pickers,
+ /for \(const model of connectedMatches\) \{\s*if \(pinnedConnectedSet\.has\(model\.id\)\) continue;/,
+ );
+ assert.match(
+ pickers,
+ /connectedMatches\s*\.filter\(\(model\) => pinnedConnectedSet\.has\(model\.id\)\)/,
+ );
+});
+
+test("connected rows join the keyboard order", () => {
+ assert.match(
+ pickers,
+ /if \(section === "connected"\) \{\s*if \(!pinnedConnectedCollapsed\) \{\s*keys\.push\(\s*\.\.\.pinnedConnectedRows\.map/,
+ );
+ // A folded group's rows leave the order too, or the arrows walk through rows off screen.
+ assert.match(
+ pickers,
+ /if \(collapsedConnectedGroups\.has\(group\.providerId\)\) continue;/,
+ );
+});
+
+test("nothing on a connected row opens local run settings", () => {
+ // ModelConfigPage holds KV cache dtype and GPU layers, which a remote model has none of.
+ assert.doesNotMatch(
+ pickers,
+ /onConfigure\(model\.id, \{\s*source: "external"/,
+ );
+ assert.match(
+ selector,
+ /function handleConfigureConnection\(providerId: string\) \{\s*setOpen\(false\);\s*useSettingsDialogStore\.getState\(\)\.openConnectionSettings\(providerId\);/,
+ );
+});
+
+test("the connection request is one-shot and tab-scoped", () => {
+ assert.match(
+ settingsStore,
+ /connectionRequested:\s*\n?\s*tab === "connections" \? state\.connectionRequested : null/,
+ );
+ // Closing the dialog drops it, or it replays on the next visit.
+ assert.match(
+ settingsStore,
+ /closeDialog: \(\) =>\s*set\(\{[^}]*connectionRequested: null,/,
+ );
+ assert.match(connectionsTab, /openProviderId=\{connectionRequested\}/);
+ assert.match(
+ connectionsTab,
+ /onOpenProviderConsumed=\{consumeConnectionRequest\}/,
+ );
+});
+
+test("the deep link waits for the provider and fires once", () => {
+ // Cached providers must not seed the form before the first backend sync.
+ assert.match(providersDialog, /\[providersReady, setProvidersReady\] = useState\(false\)/);
+ assert.match(providersDialog, /onProvidersChange\(syncedProviders\);\s*setProvidersReady\(true\);/);
+ assert.match(providersDialog, /if \(!providersReady\) return;/);
+ // Ready either way: a failed sync leaves the hydrated list as all there is, and waiting on a
+ // success the backend may never give would leave the gear opening nothing offline.
+ assert.match(
+ providersDialog,
+ /toast\.error\(`Failed to load connections: \$\{message\}`\);\s*\}\s*(\/\/[^\n]*\n\s*)*if \(isMounted\) setProvidersReady\(true\);/,
+ );
+ assert.match(providersDialog, /\[openProviderId, providers, providersReady, onOpenProviderConsumed\]/);
+ assert.match(providersDialog, /if \(!provider\) return;/);
+ assert.match(
+ providersDialog,
+ /openedProviderRef\.current = openProviderId;\s*void editProvider\(provider\);\s*onOpenProviderConsumed\?\.\(\);/,
+ );
+ // The empty-list auto-open latch stays set, as tests/connections-empty-opens-form.test.ts asks.
+ assert.doesNotMatch(
+ providersDialog,
+ /autoOpenedAddFormRef\.current = false;/,
+ );
+});
+
+test("the row menu carries the connected actions", () => {
+ for (const label of ["Model info", "Copy model ID"]) {
+ assert.match(pickers, new RegExp(`label: "${label}"`));
+ }
+ // Through the shared helper, which handles the desktop shell and falls back to execCommand.
+ // navigator.clipboard is undefined there and over plain HTTP on a LAN address, and reading
+ // .writeText off it throws where no catch on the promise can report it.
+ assert.match(
+ pickers,
+ /if \(await copyToClipboard\(providerModelId\)\) \{\s*toast\.success/,
+ );
+ assert.doesNotMatch(pickers, /navigator\.clipboard\s*\n?\s*\.?writeText/);
+ // The picked model already carries into a new chat, so a default-for-new-chats pin said nothing.
+ assert.doesNotMatch(pickers, /Use by default in new chats/);
+ // Unticking a model belongs to the connection form, which owns that list.
+ assert.doesNotMatch(pickers, /Hide from this list/);
+ // The connection belongs to the heading's gear, not to a row's menu: the row menu is about the
+ // model, and a second way in said nothing the gear above it did not.
+ assert.doesNotMatch(pickers, /label: "Connection settings",/);
+ assert.doesNotMatch(
+ pickers,
+ /onSelect: \(\) => onConfigureConnection\(model\.providerId\),/,
+ );
+});
+
+test("the name sort flattens the list without dropping a model", () => {
+ // Every heading goes, and a connection whose every model is pinned loses its group entirely,
+ // so both paths have to list from the matches rather than from the groups.
+ assert.match(pickers, /if \(connectedSort !== "name"\) return groups;/);
+ assert.match(
+ pickers,
+ /for \(const model of connectedMatches\) \{\s*if \(pinnedConnectedSet\.has\(model\.id\)\) continue;/,
+ );
+ // Such a row says which connection it came from through its logo and tooltip, pinned above.
+ // Its settings are the heading gear's, or Settings then Connections where no heading is left.
+ assert.doesNotMatch(pickers, /headless && onConfigureConnection/);
+});
+
+test("the connection's own settings hang off the group heading", () => {
+ // Beside the fold chevron, in the heading's own right-hand cluster.
+ assert.match(
+ pickers,
+ /<div className="-mr-2 flex shrink-0 items-center -space-x-0\.5">\s*\{onConfigure \?/,
+ );
+ // Hover-only, as a row's gear is, while the chevron beside it stays on screen.
+ assert.match(
+ pickers,
+ /HEADING_ACTION_CLASS,\s*"opacity-0 group-hover\/heading:opacity-100/,
+ );
+ assert.match(pickers, /className=\{HEADING_ACTION_CLASS\}\s*>\s*\{collapsed \?/);
+ // The same box and glyph the row gutter's buttons use, so the columns line up.
+ assert.match(
+ pickers,
+ /const HEADING_ACTION_CLASS =\s*"flex size-5 shrink-0 items-center justify-center rounded-md/,
+ );
+ assert.match(
+ pickers,
+ /aria-label=\{configureLabel \?\? "Connection settings"\}/,
+ );
+ assert.match(
+ pickers,
+ /onConfigure=\{\s*onConfigureConnection\s*\? \(\) => onConfigureConnection\(group\.providerId\)/,
+ );
+ // Named, since a picker can hold several connections.
+ assert.match(
+ pickers,
+ /configureLabel=\{`\$\{group\.providerName\} connection settings`\}/,
+ );
+ // The Pinned group spans connections, so there is no single one for its heading to open.
+ assert.match(
+ pickers,
+ /label="Pinned"\s*collapsed=\{pinnedConnectedCollapsed\}\s*onToggle=\{\(\) =>\s*setPinnedConnectedCollapsed/,
+ );
+});
+
+test("the gear opens the model's own settings", () => {
+ // Not a menu entry as well: the gutter already holds a settings button.
+ assert.match(
+ pickers,
+ /tooltip="Model settings"\s*onConfigure=\{\(\) =>\s*setSettingsModel\(\{/,
+ );
+ assert.doesNotMatch(pickers, /label: "Model settings/);
+});
+
+test("the connected sort and modality filter drive the list", () => {
+ assert.match(pickers, /CONNECTED_SORT_OPTIONS/);
+ assert.match(pickers, /CONNECTED_MODALITY_OPTIONS/);
+ // Sorting by name flattens the provider groups rather than leaving headings of one row each.
+ assert.match(pickers, /if \(connectedSort !== "name"\) return groups;/);
+});
+
+test("per-model prompt and cap reuse the memory Chat already keeps", () => {
+ const settingsDialog = readSrc(
+ "features/model-picker/components/model-selector/connected-model-settings-dialog.tsx",
+ );
+ // Not a second store: the editor writes into paramsByModel, which rememberParamsPerModel has
+ // been restoring per checkpoint all along.
+ assert.match(settingsDialog, /setRememberedParamsForModel\(checkpointId, \{/);
+ assert.doesNotMatch(settingsDialog, /localStorage/);
+ // Only the fields the user touched: this is a patch, and an untouched one would put whatever
+ // the dialog happened to be showing over the stored value. null is untouched, so an untouched
+ // field keeps reading the store even when hydration lands while the dialog is open.
+ assert.match(settingsDialog, /const systemPrompt = promptDraft \?\? remembered\?\.systemPrompt \?\? "";/);
+ // And a touched one is checked against the ladder as it stands at Save, since the catalogue
+ // subscription can withdraw a level while the dialog is open: a draft the model no longer
+ // offers reverts to what is stored rather than being written as a pin every resolver refuses.
+ assert.match(
+ settingsDialog,
+ /const liveEffortDraft =\s*effortDraft !== null && \(effortDraft === FOLLOW_CHAT \|\| offered\(effortDraft\)\)\s*\? effortDraft\s*: null;/,
+ );
+ assert.match(
+ settingsDialog,
+ /const effort =\s*liveEffortDraft \?\?\s*\(pinnedEffort && offered\(pinnedEffort\) \? pinnedEffort : FOLLOW_CHAT\);/,
+ );
+ // And the save is the only place a pin is read for display: the info box does not print one,
+ // so a withdrawn level has nowhere left to be claimed as applying.
+ assert.doesNotMatch(
+ readSrc(
+ "features/model-picker/components/model-selector/connected-model-info-dialog.tsx",
+ ),
+ /pinnedEffort/,
+ );
+ assert.match(
+ settingsDialog,
+ /\.\.\.\(promptDraft !== null \? \{ systemPrompt: promptDraft \} : \{\}\),/,
+ );
+ // An untouched effort writes nothing: another tab can change the pin while this is open, and a
+ // save of an unrelated field would put the value the select opened with back over it, and reset
+ // the composer's own Think level while it was at it.
+ assert.match(
+ settingsDialog,
+ /if \(liveEffortDraft !== null\) \{\s*const pinned = liveEffortDraft === FOLLOW_CHAT \? null : liveEffortDraft;\s*setModelReasoningEffort\(checkpointId, pinned\);/,
+ );
+ // Which is the listener that makes that reachable.
+ const effortStore = readSrc(
+ "features/model-picker/components/model-selector/model-reasoning-effort.ts",
+ );
+ assert.match(
+ effortStore,
+ /window\.addEventListener\("storage", \(event\) => \{/,
+ );
+ // And the chat subscribes to the active model's pin for the same reason: the normalization
+ // reads it through a getState helper, so a cross-tab change reached the dialog but not the
+ // composer. Its own effect, since rerunning that whole block would reset the pills too.
+ const chatPage = readSrc("features/chat/chat-page.tsx");
+ assert.match(
+ chatPage,
+ /const activePinnedEffort = useModelReasoningEffortStore\(\s*\(state\) => state\.effortByModel\[inferenceParams\.checkpoint\],/,
+ );
+ assert.match(
+ chatPage,
+ /if \(appliedPinnedEffort\.current === activePinnedEffort\) return;/,
+ );
+ // One reconciler for every trigger that can dislodge the pin, the catalogue included: a refresh
+ // decides which levels a stored pin is legal against, so it can make an ignored pin the valid
+ // one, and the pin effect's own guard sees no change in the stored string.
+ assert.match(
+ chatPage,
+ /reconcilePinnedReasoningEffort\(\{\s*checkpoint: inferenceParams\.checkpoint,\s*caps,\s*providerType: provider\?\.providerType,\s*\}\);\s*\}, \[activePinnedEffort,/,
+ );
+ assert.match(
+ chatPage,
+ /reasoningFieldsAfterCatalogRefresh\(useChatRuntimeStore\.getState\(\), caps\),\s*\);[\s\S]{0,220}?reconcilePinnedReasoningEffort\(\{/,
+ );
+ // Whether the level came from a pin is asked of the store, never inferred from the trigger:
+ // absence is a clear for the pin effect but not for a refresh, and it is a clear the user can
+ // also cause from the composer, where resolving afresh would undo the level just chosen.
+ assert.match(readSrc("features/chat/stores/chat-runtime-store.ts"), /if \(!pinned && !pinHoldsLiveEffort\(\)\) return;/);
+ // Switching to an unpinned model resolves from the chat's own level, not from the outgoing
+ // model's pin, which is what the live value holds.
+ assert.match(
+ chatPage,
+ /current:\s*!pinnedEffort && pinHoldsLiveEffort\(\)\s*\? \(takeEffortDisplacedByPin\(\) \?\? store\.reasoningEffort\)\s*: store\.reasoningEffort,/,
+ );
+ assert.match(
+ chatPage,
+ /current:\s*!pinnedEffort && pinHoldsLiveEffort\(\)\s*\? \(takeEffortDisplacedByPin\(\) \?\? state\.reasoningEffort\)\s*: state\.reasoningEffort,/,
+ );
+ // A pin in force owns the live level outright, whether or not it displaced anything when it
+ // was applied: pinning the level already in the store recorded nothing, and the thread opened
+ // next had its own level held back behind that pin with no record to hand back.
+ assert.match(
+ readSrc("features/chat/stores/chat-runtime-store.ts"),
+ /export function pinHoldsLiveEffort\(\): boolean \{\s*return \(\s*pinOwnsLiveReasoningEffort\(useChatRuntimeStore\.getState\(\)\) \|\|\s*effortDisplacedByPin !== null\s*\);/,
+ );
+ // And the chat's own level is recorded as the snapshot is held back, so the pin can be handed
+ // back the level of the chat that is open now rather than the one before it.
+ assert.match(
+ readSrc("features/chat/stores/chat-runtime-store.ts"),
+ /if \(key === "reasoningEffort" && pinOwnsLiveReasoningEffort\(state\)\) \{[\s\S]{0,500}?effortDisplacedByPin = value as ReasoningEffort;/,
+ );
+ // Leaving a pinned model for a local one hands the chat's level to the clamp, which narrows it
+ // to the local model's ladder and would otherwise carry the pin's level onto that model. Taken
+ // where the model has actually become resident, never where one was picked: picking an uncached
+ // model only queues a download, and the load can still abort at a lease, a confirmation or a
+ // token prompt without the model ever changing.
+ const localLoad = readSrc("features/chat/hooks/use-chat-model-runtime.ts");
+ assert.match(
+ localLoad,
+ /const existingReasoningEffort =\s*\(pinHoldsLiveEffort\(\) \? takeEffortDisplacedByPin\(\) : null\) \?\?\s*useChatRuntimeStore\.getState\(\)\.reasoningEffort;/,
+ );
+ const statusApply = readSrc("features/chat/lib/apply-inference-status-to-store.ts");
+ assert.match(
+ statusApply,
+ /const effortToClamp =\s*\(pinHoldsLiveEffort\(\) \? takeEffortDisplacedByPin\(\) : null\) \?\?\s*prevState\.reasoningEffort;/,
+ );
+ // Both clamps read the one value, so neither can keep clamping the pin's.
+ assert.match(
+ statusApply,
+ /\? clampReasoningEffortToLevels\(effortToClamp, reasoningEffortLevels\)\s*: clampLocalReasoningEffort\(effortToClamp\);/,
+ );
+ // And the pick itself takes nothing, so an abort leaves the pin applied with nothing to undo.
+ assert.doesNotMatch(chatPage, /takeEffortDisplacedByPin\(\) : null/);
+ // A number field accepts scientific notation and parseInt stops at the "e", so 1e5 was read as
+ // 1 and saved clamped to the provider minimum.
+ assert.match(
+ settingsDialog,
+ /const typedCap = Math\.round\(Number\(maxTokens\.trim\(\)\)\);/,
+ );
+ assert.doesNotMatch(settingsDialog, /Number\.parseInt\(/);
+ // The effort control is offered only where a level is actually sent. The default low/medium/high
+ // ladder is there even for a style that carries a bare thinking on/off, so gating on
+ // supportsReasoning alone let a pin be set on Kimi that no request could ever carry.
+ assert.match(
+ settingsDialog,
+ /const efforts = externalReasoningTakesEffort\(reasoning\)\s*\? reasoning\.reasoningEffortLevels\.filter\(\(level\) => level !== "none"\)\s*: \[\];/,
+ );
+ const capabilities = readSrc("features/chat/provider-capabilities.ts");
+ assert.match(
+ capabilities,
+ /export function externalReasoningTakesEffort\([\s\S]{0,120}?return caps\.supportsReasoning && caps\.reasoningStyle === "reasoning_effort";/,
+ );
+ // And nothing resolves a level for such a model, or selecting it would move the level every
+ // other model runs at on the strength of a setting it never sends.
+ assert.match(
+ capabilities,
+ /if \(!externalReasoningTakesEffort\(caps\)\) return current;/,
+ );
+ // The info box stops listing the ladder for them too: the levels are empty unless a level is
+ // what the style sends, so a bare thinking switch reports "Supported" instead of a ladder of
+ // rungs it never takes.
+ const infoDialog = readSrc(
+ "features/model-picker/components/model-selector/connected-model-info-dialog.tsx",
+ );
+ assert.match(
+ infoDialog,
+ /const takesEffort = externalReasoningTakesEffort\(reasoning\);/,
+ );
+ assert.match(
+ infoDialog,
+ /const effortLevels = takesEffort\s*\? reasoning\.reasoningEffortLevels\.filter\(\(level\) => level !== "none"\)\s*: \[\];/,
+ );
+ // Blank or junk is an absence, not a zero the request would then send as the cap.
+ assert.match(
+ settingsDialog,
+ /capDraft !== null && Number\.isFinite\(typedCap\) && typedCap > 0\s*\? \{ maxTokens: clampCap\(typedCap\) \}\s*: \{\}/,
+ );
+ // And it says so when the setting that restores them is switched off.
+ assert.match(settingsDialog, /"Remember settings per model" is off/);
+ // No example prompt: a placeholder reads as a value that is already set.
+ assert.doesNotMatch(settingsDialog, /placeholder=/);
+ // The cap field always holds a number. A blank would have to mean "forget the cap", and
+ // neither merge can express that: the patch merges per key and the server deep-merges the
+ // settings row, so an omitted key keeps the old value and the clear goes nowhere.
+ assert.match(
+ settingsDialog,
+ /capDraft \?\? String\(clampCap\(remembered\?\.maxTokens \?\? chatMaxTokens\)\)/,
+ );
+ assert.doesNotMatch(settingsDialog, /blank follows the connection/);
+});
+
+test("the cap offers the bounds every request is clamped to", () => {
+ const settingsDialog = readSrc(
+ "features/model-picker/components/model-selector/connected-model-settings-dialog.tsx",
+ );
+ const adapter = readSrc("features/chat/api/chat-adapter.ts");
+ const sheet = readSrc("features/chat/chat-settings-sheet.tsx");
+ // The same two helpers the chat's own Max Tokens control bounds itself with, and the adapter
+ // clamps every outbound request to, so a stored cap cannot differ from the one actually sent.
+ assert.match(sheet, /getExternalMaxOutputTokens\(/);
+ assert.match(adapter, /getExternalMinOutputTokens\(externalProvider\?\.providerType\),/);
+ assert.match(
+ settingsDialog,
+ /const minCap = getExternalMinOutputTokens\(providerType\);/,
+ );
+ assert.match(
+ settingsDialog,
+ /const maxCap = getExternalMaxOutputTokens\(\s*providerType,\s*modelId,\s*connectionMaxOutputTokens,\s*\)/,
+ );
+ // Clamped on save, not just hinted by the input's attributes.
+ assert.match(
+ settingsDialog,
+ /const clampCap = \(value: number\) =>\s*Math\.min\(Math\.max\(value, minCap\), maxCap\);/,
+ );
+ // The connection's own cap lowers the model's documented one, so it has to reach the editor.
+ assert.match(
+ pickers,
+ /provider\.maxOutputTokens \?\? null,/,
+ );
+ // And an OpenRouter cap arrives with the live catalogue, after this has rendered.
+ assert.match(
+ settingsDialog,
+ /useSyncExternalStore\(subscribeModelCatalog, modelCatalogVersion\);/,
+ );
+});
+
+test("live effort edits use the shared runtime action", () => {
+ const dialog = readSrc(
+ "features/model-picker/components/model-selector/connected-model-settings-dialog.tsx",
+ );
+ assert.match(dialog, /reconcilePinnedReasoningEffort\(\{/);
+ assert.doesNotMatch(dialog, /useChatRuntimeStore\.setState/);
+ assert.doesNotMatch(dialog, /state\.setReasoningEffort/);
+});
+
+test("a row with no heading over it still names its connection", () => {
+ // Pinned rows and the name-sorted flat list have no provider heading, and two connections can
+ // serve one model id, so the tooltip carries the connection name those rows have nowhere else.
+ assert.match(pickers, /renderConnectedModelRow\(model, true, true\)/);
+ assert.match(pickers, /renderConnectedModelRow\(model, false, !headed\)/);
+ assert.match(pickers, /<span className="block text-ui-10 mt-1">\s*\{model\.providerName\}/);
+ // No logo in the leading slot: down the pinned group it read as a second glyph column, and a
+ // row is there to carry its name. So nothing goes in the slot at all now.
+ assert.doesNotMatch(pickers, /leadingSlot/);
+ assert.match(
+ pickers,
+ /const leading = formatDot \? <FormatTag \{\.\.\.formatDot\} \/> : null;/,
+ );
+ // The tooltip opens to the right. This panel is docked at the window's left edge, so a row's
+ // own left edge is ~30px in and a tooltip opening that way ran off screen.
+ assert.match(
+ pickers,
+ /<TooltipContent\s*side="right"\s*className="tooltip-compact max-w-\[15rem\] break-words"/,
+ );
+});
+
+test("the store write merges per key and reaches the live params", () => {
+ const runtime = readSrc("features/chat/stores/chat-runtime-store.ts");
+ assert.match(
+ runtime,
+ /saveSettingsPatch\(\{ inferenceParamsByModel: \{ \[modelId\]: patch \} \}\)/,
+ );
+ // Editing the model that is loaded has to land now, not on the next switch back.
+ assert.match(runtime, /const live = state\.params\.checkpoint === modelId;/);
+ // A save while the initial settings request is still out has to survive it, so the write is not
+ // gated on hydration: the patch can only set the keys it names and the server merges per key.
+ assert.match(
+ runtime,
+ /\/\/ \/api\/chat\/settings request was out\.\s*saveSettingsPatch\(/,
+ );
+ // Held as a patch, not a row. locallyRememberedModels is overlaid wholesale, which would
+ // replace the server's entry with one naming only the keys the dialog touched and drop that
+ // model's temperature, top-p and seed.
+ assert.match(
+ runtime,
+ /if \(!state\.settingsHydrated\) \{\s*modelParamEditsBeforeHydration\.set\(modelId, \{/,
+ );
+ assert.match(
+ runtime,
+ /for \(const \[modelId, patch\] of modelParamEditsBeforeHydration\) \{\s*hydrated\[modelId\] = \{ \.\.\.hydrated\[modelId\], \.\.\.patch \};/,
+ );
+ // Complete once merged, so a later response takes the wholesale overlay.
+ assert.match(
+ runtime,
+ /locallyRememberedModels\.add\(modelId\);\s*\}\s*modelParamEditsBeforeHydration\.clear\(\);/,
+ );
+ // And the live keys are fenced the way a slider edit fences its own, or the response in flight
+ // puts the global set back over them.
+ assert.match(
+ runtime,
+ /if \(liveChanged\) getChangedInferenceParams\(liveParams, state\.params\);/,
+ );
+ // systemPrompt is one of the chat's own keys, so the live overlay goes through the restore a
+ // switch uses: the open chat outranks the model it is running, and writing past that would
+ // send the model's prompt for the rest of the chat and store it as the chat's own next.
+ assert.match(
+ runtime,
+ /const liveParams = live\s*\? restoreThreadScopedParams\(\{ \.\.\.state\.params, \.\.\.patch \}\)\s*: null;/,
+ );
+ // The epoch as setParams advances it: a queued prompt or a paste still reading its file
+ // captured the old prompt and cap, and nothing else would tell it they had moved.
+ assert.match(
+ runtime,
+ /const liveChanged =\s*liveParams !== null &&\s*shouldAdvanceQueuedSettingsEpoch\(state\.params, liveParams\);/,
+ );
+ assert.match(
+ runtime,
+ /liveChanged\s*\? \{\s*params: liveParams,\s*queuedSettingsEpoch: state\.queuedSettingsEpoch \+ 1,/,
+ );
+});
+
+test("a pinned reasoning effort wins, unless the catalogue withdrew it", () => {
+ const effort = readSrc(
+ "features/model-picker/components/model-selector/model-reasoning-effort.ts",
+ );
+ // Asking for a level the provider rejects fails the request, so an unavailable pin falls back.
+ assert.match(
+ effort,
+ /if \(!effort \|\| \(allowed && !allowed\.includes\(effort\)\)\) return null;/,
+ );
+ const caps = readSrc("features/chat/provider-capabilities.ts");
+ // Ahead of the catalogue and the per-provider defaults: the user set this one deliberately.
+ assert.match(
+ caps,
+ /if \(pinned && levels\.includes\(pinned as ReasoningEffortLevel\)\) \{\s*return pinned as ReasoningEffortLevel;/,
+ );
+ assert.match(
+ caps,
+ /if \(caps\.defaultEffort && levels\.includes\(caps\.defaultEffort\)\) \{/,
+ );
+ // One resolver for every caller, so a reload and a resync cannot answer differently from a
+ // switch. The pin used to be read by the switch alone.
+ const chatPage = readSrc("features/chat/chat-page.tsx");
+ assert.match(
+ chatPage,
+ /externalReasoningTakesEffort\(reasoningCaps\)\s*\? pinnedReasoningEffort\(value, effortLevels\)/,
+ );
+ assert.match(
+ chatPage,
+ /externalReasoningTakesEffort\(reasoningCaps\)\s*\? pinnedReasoningEffort\(inferenceParams\.checkpoint, effortLevels\)/,
+ );
+ // Both record the level the pin takes the place of, so clearing it can put that back.
+ assert.match(
+ chatPage,
+ /if \(pinnedEffort && nextReasoningEffort !== store\.reasoningEffort\) \{\s*noteEffortDisplacedByPin\(store\.reasoningEffort\);/,
+ );
+ assert.match(
+ chatPage,
+ /if \(pinnedEffort && nextReasoningEffort !== state\.reasoningEffort\) \{\s*noteEffortDisplacedByPin\(state\.reasoningEffort\);/,
+ );
+ // The normalization effect reruns on reload and on every provider resync, and it is the one
+ // that used to put the provider default back over the pin.
+ assert.match(
+ chatPage,
+ /\}, \[externalProvidersForChat, inferenceParams\.checkpoint, settingsHydrated\]\);/,
+ );
+ assert.doesNotMatch(chatPage, /const anthropicTopEffort =/);
+});
+
+test("the info box is a dialog, so it is wide enough and closable", () => {
+ const infoDialog = readSrc(
+ "features/model-picker/components/model-selector/connected-model-info-dialog.tsx",
+ );
+ // AlertDialog caps at max-w-md and asks a question; this reports, and needs Dialog's close X.
+ assert.match(infoDialog, /<DialogContent className="sm:max-w-xl">/);
+ assert.doesNotMatch(infoDialog, /AlertDialog/);
+ // An id has no spaces to wrap at, so break-words would let it run past the edge.
+ assert.match(infoDialog, /className="break-all font-mono"/);
+ // It reports the model, not the settings. The gear beside it edits those three fields and shows
+ // what they hold, so repeating them here was a second, read-only copy of that dialog.
+ assert.doesNotMatch(infoDialog, /own settings/);
+ assert.doesNotMatch(infoDialog, /paramsByModel/);
+ assert.doesNotMatch(infoDialog, /model-reasoning-effort/);
+ // So it needs no id to key those memories on either.
+ assert.doesNotMatch(infoDialog, /checkpointId/);
+ assert.doesNotMatch(pickers, /checkpointId=\{infoModel\.model\.id\}/);
+});
+
+test("a row's name starts where its heading's label does", () => {
+ // ml-4 plus pl-3.5 is 30px, which is the heading's px-2.5 + a size-3.5 icon + gap-1.5: where
+ // its label starts. The two have to keep summing to 30, since the second of them is the pill's
+ // own inset: widening it moves the pill's left edge, not the name off that label. 14px of pill
+ // now precedes the name, against the 23.5px the reserved leading slot used to leave empty.
+ assert.match(
+ pickers,
+ /cn\(downloadedRowShellClassName\(isSelected\), "ml-4"\)/,
+ );
+ assert.match(
+ pickers,
+ /className=\{cn\(downloadedRowButtonClassName, "pl-3\.5"\)\}/,
+ );
+ // Nothing precedes the name in the row, so the slot itself goes: an empty one is what put the
+ // pill's left edge that far from the name.
+ assert.doesNotMatch(pickers, /reserveLeadingSlot=\{true\}/);
+ // And the heading matches ListLabel, which is what the On Device sections use, so the first
+ // one sits at the same height on both tabs with its label the same distance from its icon.
+ assert.match(
+ pickers,
+ /group\/heading flex items-center justify-between gap-1 px-2\.5 pb-1 pt-3/,
+ );
+ assert.match(
+ pickers,
+ /<span className="flex min-w-0 items-center gap-1\.5 text-ui-10 font-semibold/,
+ );
+ const listLabel = /className=\{cn\(\s*"flex items-center justify-between gap-1 px-2\.5 pb-1",\s*divider \? "mt-3 border-t border-border\/50 pt-3" : "pt-3",/;
+ assert.match(pickers, listLabel);
+ assert.match(
+ pickers,
+ /<span className="flex items-center gap-1\.5 text-ui-10 font-semibold/,
+ );
+});
+
+test("reasoning is read through the resolver the composer uses", () => {
+ const infoDialog = readSrc(
+ "features/model-picker/components/model-selector/connected-model-info-dialog.tsx",
+ );
+ const settingsDialog = readSrc(
+ "features/model-picker/components/model-selector/connected-model-settings-dialog.tsx",
+ );
+ // Reading the catalogue directly reported every Codex model as unpublished while the composer
+ // had it thinking: `openai_codex` is not a namespace models.dev publishes.
+ for (const source of [infoDialog, settingsDialog]) {
+ assert.match(
+ source,
+ /getExternalReasoningCapabilities\(providerType, modelId, \{\s*isReasoningProvider,\s*baseUrl,\s*\}\)/,
+ );
+ // "none" is the off switch, not a level on offer.
+ assert.match(source, /\(level\) => level !== "none"/);
+ }
+ assert.doesNotMatch(settingsDialog, /entry\?\.reasoning \? entry\.efforts/);
+ // And a self-hosted endpoint's only reasoning signal is the flag on its connection.
+ assert.match(pickers, /provider\.isReasoningModel === true,/);
+});
+
+test("a Codex connection resolves against OpenAI's catalogue", async () => {
+ // Behaviour rather than the shape of the alias table: a null entry is what the info panel
+ // renders as "Not published", which is the whole symptom, and the alias is applied at the
+ // snapshot lookup so every caller of it inherits this.
+ const { resolveModelCatalogEntry } = await import("../src/features/chat/model-catalog.ts");
+ const viaOpenAI = resolveModelCatalogEntry("openai", "gpt-4-turbo");
+ assert.ok(viaOpenAI !== null, "the fixture model must be in the bundled snapshot");
+ assert.deepEqual(resolveModelCatalogEntry("openai_codex", "gpt-4-turbo"), viaOpenAI);
+});
+
+test("the published context window reaches the row and the info box", () => {
+ const catalog = readSrc("features/chat/model-catalog.ts");
+ const snapshot = readSrc("features/chat/model-catalog-snapshot.ts");
+ const infoDialog = readSrc(
+ "features/model-picker/components/model-selector/connected-model-info-dialog.tsx",
+ );
+ assert.match(snapshot, /context\?: number;/);
+ // models.dev publishes it on the snapshot only, so a live entry backfills from there rather
+ // than reporting a model that has one as having none.
+ assert.match(
+ catalog,
+ /if \(entry\.contextLength != null \|\| !snapshot\) return entry;/,
+ );
+ assert.match(
+ catalog,
+ /\.find\(\(context\) => typeof context === "number" && context > 0\)/,
+ );
+ // The info box alone, with the exact count. No chip on the row: the name is what a row is for,
+ // and a window on every one of them read as noise rather than as the figure to compare.
+ assert.doesNotMatch(pickers, /ctx`\]/);
+ assert.doesNotMatch(marks, /formatContextLength/);
+ assert.doesNotMatch(marks, /contextLength/);
+ assert.match(infoDialog, /<Field label="Context window">/);
+ assert.match(infoDialog, /tokens\(entry\.contextLength\)/);
+});
+
+test("a served catalogue cannot take away a context window it has no field for", async () => {
+ // Behaviour, not source text. A regex over these two files passes just as happily with the
+ // line present in code that never runs, and fails on a rename that changes nothing; the point
+ // is what a user reads in the info panel, so drive the resolver and read the window back.
+ const { resolveModelCatalogEntry, setModelsDevCatalog } = await import(
+ "../src/features/chat/model-catalog.ts"
+ );
+ const bundled = resolveModelCatalogEntry("openai", "gpt-4-turbo")?.contextLength;
+ assert.ok(
+ typeof bundled === "number" && bundled > 0,
+ "the bundled snapshot must publish a window for this model, else the rest proves nothing",
+ );
+
+ // A served entry that carries one wins, as it does for every other field.
+ setModelsDevCatalog({
+ fetched_at: Date.now(),
+ providers: { openai: { "gpt-4-turbo": { input: ["text"], context: 999_999 } } },
+ } as never);
+ assert.equal(resolveModelCatalogEntry("openai", "gpt-4-turbo")?.contextLength, 999_999);
+
+ // THE CASE: a backend older than the field, or its cached payload inside the day-long TTL,
+ // serves the model with no context at all. Before the rescue this replaced the bundled entry
+ // wholesale and every model models.dev covers read as publishing no window.
+ setModelsDevCatalog({
+ fetched_at: Date.now(),
+ providers: { openai: { "gpt-4-turbo": { input: ["text"] } } },
+ } as never);
+ assert.equal(resolveModelCatalogEntry("openai", "gpt-4-turbo")?.contextLength, bundled);
+
+ setModelsDevCatalog({ fetched_at: Date.now(), providers: {} } as never);
+});
diff --git a/studio/frontend/tests/helpers/kit.ts b/studio/frontend/tests/helpers/kit.ts
index cee4fba2376..ce64025d8be 100644
--- a/studio/frontend/tests/helpers/kit.ts
+++ b/studio/frontend/tests/helpers/kit.ts
@@ -53,11 +53,13 @@ export type StorageFake = {
*/
export function installLocalStorageFake(): {
store: Map<string, string>;
- storage: StorageFake;
+ storage: StorageFake & Pick<Storage, "length" | "key">;
fireWindowEvent: (type: string, event: unknown) => number;
} {
const store = new Map<string, string>();
- const storage: StorageFake = {
+ const storage: StorageFake & Pick<Storage, "length" | "key"> = {
+ get length() { return store.size; },
+ key: (index) => [...store.keys()][index] ?? null,
getItem: (key: string) => store.get(key) ?? null,
setItem: (key: string, value: string) => {
store.set(key, value);
diff --git a/studio/frontend/tests/model-reasoning-preferences.test.ts b/studio/frontend/tests/model-reasoning-preferences.test.ts
new file mode 100644
index 00000000000..164c58e7dc0
--- /dev/null
+++ b/studio/frontend/tests/model-reasoning-preferences.test.ts
@@ -0,0 +1,306 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import assert from "node:assert/strict";
+import { register } from "node:module";
+import test from "node:test";
+import type { ExternalReasoningCapabilities } from "../src/features/chat/provider-capabilities.ts";
+import { installLocalStorageFake } from "./helpers/kit.ts";
+
+const { store: storageData, storage, fireWindowEvent } = installLocalStorageFake();
+register("./thread-sampling-resolver.mjs", import.meta.url);
+const { settingsHttp } = await import("./helpers/store-stubs/settings-http.ts");
+const { threadRows } = await import("./helpers/store-stubs/chat-history-storage.ts");
+const {
+ createModelReasoningEffortStore,
+ useModelReasoningEffortStore,
+ pinnedReasoningEffort,
+} = await import("../src/features/model-picker/components/model-selector/model-reasoning-effort.ts");
+const { resolveExternalReasoningEffort } = await import("../src/features/chat/provider-capabilities.ts");
+
+const KEY = "unsloth_model_reasoning_effort";
+const MODEL = "external::review::reasoning-model";
+const THREAD = "review-thread";
+const CAPS: ExternalReasoningCapabilities = {
+ supportsReasoning: true,
+ reasoningStyle: "reasoning_effort",
+ reasoningAlwaysOn: false,
+ supportsReasoningOff: true,
+ reasoningEffortLevels: ["low", "medium", "high"],
+};
+type RuntimeModule = typeof import("../src/features/chat/stores/chat-runtime-store.ts");
+const runtimeUrl = new URL("../src/features/chat/stores/chat-runtime-store.ts", import.meta.url).href;
+
+async function boot(scenario: string, paired: boolean | null = true): Promise<RuntimeModule> {
+ storageData.clear();
+ storageData.set("unsloth_chat_settings_imported_to_studio_db", "true");
+ useModelReasoningEffortStore.getState().syncFromStorage();
+ settingsHttp.settings = { rememberParamsPerModel: false, reasoningEffort: "medium" };
+ settingsHttp.puts.length = 0;
+ threadRows.reset();
+ const runtime = await import(`${runtimeUrl}?scenario=${scenario}`) as RuntimeModule;
+ const store = runtime.useChatRuntimeStore;
+ await store.getState().hydratePersistedSettings();
+ store.getState().setCheckpoint(MODEL, null);
+ store.setState({ ...CAPS, reasoningEffort: "medium" });
+ if (paired !== null) store.getState().setActiveThreadId(THREAD);
+ if (paired) store.getState().applyThreadScopedSettings(THREAD, { reasoningEffort: "medium" });
+ else if (paired === false) runtime.beginThreadScopedPairing(THREAD);
+ settingsHttp.puts.length = 0;
+ return runtime;
+}
+
+function remotePin(runtime: RuntimeModule, effort: string | null): void {
+ createModelReasoningEffortStore(storage).getState().setModelReasoningEffort(MODEL, effort);
+ fireWindowEvent("storage", { key: `${KEY}::${MODEL}`, storageArea: storage });
+ runtime.reconcilePinnedReasoningEffort({ checkpoint: MODEL, caps: CAPS, providerType: "openai" });
+}
+
+test("independent pin edits survive delayed storage events", () => {
+ storageData.clear();
+ const a = createModelReasoningEffortStore(storage);
+ const b = createModelReasoningEffortStore(storage);
+ a.getState().setModelReasoningEffort("model-a", "low");
+ b.getState().setModelReasoningEffort("model-b", "high");
+ a.getState().syncFromStorage();
+ b.getState().syncFromStorage();
+ const expected = { "model-a": "low", "model-b": "high" };
+ assert.deepEqual(a.getState().effortByModel, expected);
+ assert.deepEqual(b.getState().effortByModel, expected);
+ assert.deepEqual(createModelReasoningEffortStore(storage).getState().effortByModel, expected);
+});
+
+test("clearing an unrelated pin does not erase a delayed edit", () => {
+ storageData.clear();
+ storageData.set(KEY, JSON.stringify({ "model-b": "medium" }));
+ const a = createModelReasoningEffortStore(storage);
+ const b = createModelReasoningEffortStore(storage);
+ a.getState().setModelReasoningEffort("model-a", "high");
+ b.getState().setModelReasoningEffort("model-b", null);
+ assert.deepEqual(createModelReasoningEffortStore(storage).getState().effortByModel, { "model-a": "high" });
+});
+
+test("legacy pins load without writes and cleared pins cannot return", () => {
+ storageData.clear();
+ const legacy = JSON.stringify({ a: "low", b: "medium" });
+ storageData.set(KEY, legacy);
+ const a = createModelReasoningEffortStore(storage);
+ assert.equal(storageData.size, 1);
+ assert.deepEqual(a.getState().effortByModel, { a: "low", b: "medium" });
+ a.getState().setModelReasoningEffort("a", null);
+ const b = createModelReasoningEffortStore(storage);
+ assert.deepEqual(b.getState().effortByModel, { b: "medium" });
+ b.getState().setModelReasoningEffort("b", "high");
+ assert.equal(storageData.get(KEY), legacy);
+ assert.deepEqual(createModelReasoningEffortStore(storage).getState().effortByModel, { b: "high" });
+});
+
+test("same-model edits use the last persisted write", () => {
+ storageData.clear();
+ const a = createModelReasoningEffortStore(storage);
+ const b = createModelReasoningEffortStore(storage);
+ a.getState().setModelReasoningEffort("a", "low");
+ b.getState().setModelReasoningEffort("a", "high");
+ a.getState().syncFromStorage();
+ assert.equal(a.getState().effortByModel.a, "high");
+ a.getState().setModelReasoningEffort("a", null);
+ b.getState().syncFromStorage();
+ assert.deepEqual(b.getState().effortByModel, {});
+});
+
+test("malformed legacy storage does not hide newer pins", () => {
+ storageData.clear();
+ storageData.set(KEY, "invalid json");
+ storageData.set(`${KEY}::a`, "high");
+ assert.deepEqual(createModelReasoningEffortStore(storage).getState().effortByModel, { a: "high" });
+});
+
+test("unavailable storage keeps edits in memory", () => {
+ const store = createModelReasoningEffortStore(null);
+ store.getState().setModelReasoningEffort("a", "low");
+ assert.equal(store.getState().effortByModel.a, "low");
+});
+
+test("an empty offered ladder rejects stored pins", () => {
+ useModelReasoningEffortStore.setState({ effortByModel: { a: "high" } });
+ assert.equal(pinnedReasoningEffort("a", []), null);
+ assert.equal(pinnedReasoningEffort("a", ["high"]), "high");
+});
+
+test("restoration preserves a valid chat preference without changing selection defaults", () => {
+ for (const [providerType, expected] of [["openai", "high"], ["anthropic", "high"], ["gemini", "medium"]]) {
+ assert.equal(resolveExternalReasoningEffort({ caps: CAPS, providerType, current: "low" }), expected);
+ assert.equal(resolveExternalReasoningEffort({ caps: CAPS, providerType, current: "low", restore: true }), "low");
+ assert.equal(resolveExternalReasoningEffort({ caps: { ...CAPS, defaultEffort: "high" }, providerType, current: "low", restore: true }), "low");
+ }
+});
+
+test("restoration clamps unavailable levels and ignores toggle-only pins", () => {
+ assert.equal(resolveExternalReasoningEffort({ caps: { ...CAPS, reasoningEffortLevels: ["low", "high"] }, providerType: "openai", current: "medium", restore: true }), "low");
+ assert.equal(resolveExternalReasoningEffort({ caps: { ...CAPS, reasoningStyle: "enable_thinking" }, providerType: "openai", current: "low", pinned: "high", restore: true }), "low");
+});
+
+test("a pin arriving during the debounce does not become the thread preference", async (t) => {
+ const runtime = await boot("pending-edit");
+ const store = runtime.useChatRuntimeStore;
+ t.mock.timers.enable({ apis: ["setTimeout"] });
+ store.getState().setReasoningEffort("low");
+ t.mock.timers.tick(100);
+ remotePin(runtime, "high");
+ assert.equal(store.getState().reasoningEffort, "high");
+ t.mock.timers.tick(300);
+ await runtime.awaitStartedThreadScopedSettingsWrites();
+ assert.equal(threadRows.rows.get(THREAD)?.reasoningEffort, "low");
+ assert.equal(store.getState().reasoningEffort, "high");
+ assert.equal(settingsHttp.puts.some((patch) => "reasoningEffort" in patch), false);
+});
+
+test("an ordinary pending edit persists without a pin", async (t) => {
+ const runtime = await boot("ordinary-edit");
+ t.mock.timers.enable({ apis: ["setTimeout"] });
+ runtime.useChatRuntimeStore.getState().setReasoningEffort("low");
+ t.mock.timers.tick(400);
+ await runtime.awaitStartedThreadScopedSettingsWrites();
+ assert.equal(threadRows.rows.get(THREAD)?.reasoningEffort, "low");
+});
+
+test("clearing a pin before the debounce restores the pending preference", async (t) => {
+ const runtime = await boot("clear-before-save");
+ const store = runtime.useChatRuntimeStore;
+ t.mock.timers.enable({ apis: ["setTimeout"] });
+ store.getState().setReasoningEffort("low");
+ remotePin(runtime, "high");
+ remotePin(runtime, null);
+ assert.equal(store.getState().reasoningEffort, "low");
+ t.mock.timers.tick(400);
+ await runtime.awaitStartedThreadScopedSettingsWrites();
+ assert.equal(threadRows.rows.get(THREAD)?.reasoningEffort, "low");
+});
+
+test("reconciliation advances the queue epoch only when effort changes", async () => {
+ const runtime = await boot("queue-epoch");
+ const store = runtime.useChatRuntimeStore;
+ const initial = store.getState().queuedSettingsEpoch;
+ remotePin(runtime, "high");
+ assert.equal(store.getState().queuedSettingsEpoch, initial + 1);
+ remotePin(runtime, "high");
+ assert.equal(store.getState().queuedSettingsEpoch, initial + 1);
+ remotePin(runtime, null);
+ assert.equal(store.getState().reasoningEffort, "medium");
+ assert.equal(store.getState().queuedSettingsEpoch, initial + 2);
+});
+
+test("an incoming thread keeps its own effort beneath an existing pin", async () => {
+ const runtime = await boot("switch-thread");
+ const store = runtime.useChatRuntimeStore;
+ remotePin(runtime, "high");
+ store.getState().setActiveThreadId("other-thread");
+ store.getState().applyThreadScopedSettings("other-thread", { reasoningEffort: "low" });
+ assert.equal(store.getState().reasoningEffort, "high");
+ remotePin(runtime, null);
+ assert.equal(store.getState().reasoningEffort, "low");
+});
+
+test("held effort edits survive a pin arriving before pairing completes", async (t) => {
+ const runtime = await boot("held-edit", false);
+ const store = runtime.useChatRuntimeStore;
+ t.mock.timers.enable({ apis: ["setTimeout"] });
+ store.getState().setReasoningEffort("low");
+ remotePin(runtime, "high");
+ store.getState().applyThreadScopedSettings(THREAD, { reasoningEffort: "medium" });
+ t.mock.timers.tick(400);
+ await runtime.awaitStartedThreadScopedSettingsWrites();
+ assert.equal(threadRows.rows.get(THREAD)?.reasoningEffort, "low");
+ assert.equal(store.getState().reasoningEffort, "high");
+ store.getState().setActiveThreadId(null);
+ store.getState().applyThreadScopedSettings(null, null);
+ remotePin(runtime, null);
+ assert.equal(store.getState().reasoningEffort, "medium");
+});
+
+test("leaving mid-pairing persists the requested effort rather than the pin", async () => {
+ const runtime = await boot("held-leave", false);
+ runtime.useChatRuntimeStore.getState().setReasoningEffort("low");
+ remotePin(runtime, "high");
+ runtime.useChatRuntimeStore.getState().setActiveThreadId("other-thread");
+ await runtime.commitHeldThreadScopedEditsToTheirThread();
+ assert.equal(threadRows.rows.get(THREAD)?.reasoningEffort, "low");
+});
+
+test("a capability downgrade releases an effort pin", async () => {
+ const runtime = await boot("toggle-only");
+ const store = runtime.useChatRuntimeStore;
+ remotePin(runtime, "high");
+ const caps = { ...CAPS, reasoningStyle: "enable_thinking" as const };
+ store.setState(caps);
+ runtime.reconcilePinnedReasoningEffort({ checkpoint: MODEL, caps, providerType: "openai" });
+ assert.equal(store.getState().reasoningEffort, "medium");
+ assert.equal(runtime.pinHoldsLiveEffort(), false);
+});
+
+test("a global effort edit survives pinning after leaving a saved chat", async () => {
+ const runtime = await boot("global-after-thread");
+ const store = runtime.useChatRuntimeStore;
+ store.getState().setActiveThreadId(null);
+ store.getState().applyThreadScopedSettings(null, null);
+ store.getState().setReasoningEffort("low");
+ remotePin(runtime, "high");
+ remotePin(runtime, null);
+ assert.equal(store.getState().reasoningEffort, "low");
+});
+
+test("reselecting a pinned level preserves the unsent chat's preference", async () => {
+ const runtime = await boot("reselect-pin", null);
+ const store = runtime.useChatRuntimeStore;
+ store.getState().setReasoningEffort("low");
+ remotePin(runtime, "high");
+ store.getState().setReasoningEffort("high");
+ assert.equal(pinnedReasoningEffort(MODEL, CAPS.reasoningEffortLevels), "high");
+ remotePin(runtime, null);
+ assert.equal(store.getState().reasoningEffort, "low");
+ assert.equal(runtime.pinHoldsLiveEffort(), false);
+});
+
+test("choosing a different effort clears the pin and keeps the new preference", async () => {
+ const runtime = await boot("replace-pin", null);
+ const store = runtime.useChatRuntimeStore;
+ remotePin(runtime, "high");
+ store.getState().setReasoningEffort("low");
+ assert.equal(pinnedReasoningEffort(MODEL, CAPS.reasoningEffortLevels), null);
+ remotePin(runtime, null);
+ assert.equal(store.getState().reasoningEffort, "low");
+ assert.equal(runtime.pinHoldsLiveEffort(), false);
+});
+
+test("eject restores effort before a snapshotless chat captures its defaults", async () => {
+ const runtime = await boot("eject-unsent", null);
+ const store = runtime.useChatRuntimeStore;
+ store.getState().setReasoningEffort("low");
+ remotePin(runtime, "high");
+ store.getState().clearCheckpoint();
+ assert.equal(store.getState().params.checkpoint, "");
+ assert.equal(store.getState().reasoningEffort, "low");
+ assert.equal(runtime.pinHoldsLiveEffort(), false);
+ assert.equal(pinnedReasoningEffort(MODEL, CAPS.reasoningEffortLevels), "high");
+ runtime.beginThreadScopedPairing(THREAD);
+ store.getState().applyThreadScopedSettings(null, null);
+ store.getState().setActiveThreadId(THREAD);
+ store.getState().applyThreadScopedSettings(THREAD, null);
+ assert.equal(runtime.threadScopedOverride("reasoningEffort"), "low");
+ store.getState().clearCheckpoint();
+ assert.equal(store.getState().reasoningEffort, "low");
+});
+
+test("eject restores the active thread's effort without persisting the pin", async (t) => {
+ const runtime = await boot("eject-thread");
+ const store = runtime.useChatRuntimeStore;
+ t.mock.timers.enable({ apis: ["setTimeout"] });
+ store.getState().setReasoningEffort("low");
+ remotePin(runtime, "high");
+ store.getState().clearCheckpoint();
+ assert.equal(store.getState().reasoningEffort, "low");
+ t.mock.timers.tick(400);
+ await runtime.awaitStartedThreadScopedSettingsWrites();
+ assert.equal(threadRows.rows.get(THREAD)?.reasoningEffort, "low");
+ assert.equal(settingsHttp.puts.some((patch) => "reasoningEffort" in patch), false);
+});
diff --git a/studio/frontend/tests/pinned-connected-models-reorder.test.ts b/studio/frontend/tests/pinned-connected-models-reorder.test.ts
new file mode 100644
index 00000000000..53a26d549da
--- /dev/null
+++ b/studio/frontend/tests/pinned-connected-models-reorder.test.ts
@@ -0,0 +1,93 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import assert from "node:assert/strict";
+import test from "node:test";
+import { installLocalStorageFake } from "./helpers/kit.ts";
+
+const { store: storage, fireWindowEvent } = installLocalStorageFake();
+const { usePinnedConnectedModelsStore: pins } = await import(
+ "../src/features/model-picker/components/model-selector/pinned-connected-models.ts"
+);
+const KEY = "unsloth_pinned_connected_models";
+const [A, B, C, D] = ["a", "b", "c", "d"].map((id) => `external::connection::${id}`);
+
+function reset(order = [A, B]) {
+ pins.getState().endPinnedConnectedDrag(false);
+ storage.clear();
+ storage.set(KEY, JSON.stringify(order));
+ pins.setState({ pinned: order });
+}
+
+function externalWrite(order: string[], notify = true) {
+ storage.set(KEY, JSON.stringify(order));
+ if (notify) assert.equal(fireWindowEvent("storage", { key: KEY }), 1);
+}
+
+function storedOrder(): string[] {
+ return JSON.parse(storage.get(KEY)!);
+}
+
+function drag() {
+ pins.getState().beginPinnedConnectedDrag();
+ pins.getState().movePinnedConnected(A, B);
+}
+
+test("ordinary drags persist only on drop", () => {
+ reset();
+ drag();
+ assert.deepEqual(pins.getState().pinned, [B, A]);
+ assert.deepEqual(storedOrder(), [A, B]);
+ pins.getState().endPinnedConnectedDrag(true);
+ pins.getState().endPinnedConnectedDrag(false);
+ assert.deepEqual(storedOrder(), [B, A]);
+ assert.deepEqual(pins.getState().pinned, [B, A]);
+});
+
+for (const notify of [true, false]) {
+ test(`drop keeps the dragged order and remote additions, storage event=${notify}`, () => {
+ reset();
+ drag();
+ externalWrite([C, A, B], notify);
+ assert.deepEqual(pins.getState().pinned, [B, A]);
+ pins.getState().endPinnedConnectedDrag(true);
+ assert.deepEqual(pins.getState().pinned, [C, B, A]);
+ assert.deepEqual(storedOrder(), [C, B, A]);
+ });
+}
+
+test("drop respects remote removals and the latest additions", () => {
+ reset([A, B, C]);
+ drag();
+ externalWrite([D, A, B, C]);
+ externalWrite([D, A, B]);
+ pins.getState().endPinnedConnectedDrag(true);
+ assert.deepEqual(pins.getState().pinned, [D, B, A]);
+ assert.deepEqual(storedOrder(), [D, B, A]);
+});
+
+for (const notify of [true, false]) {
+ test(`cancel adopts remote changes without persisting the drag, storage event=${notify}`, () => {
+ reset();
+ drag();
+ externalWrite([C, A], notify);
+ pins.getState().endPinnedConnectedDrag(false);
+ assert.deepEqual(pins.getState().pinned, [C, A]);
+ assert.deepEqual(storedOrder(), [C, A]);
+ });
+}
+
+test("a drag without a move adopts the remote order", () => {
+ reset([A, B, C]);
+ pins.getState().beginPinnedConnectedDrag();
+ externalWrite([C, B, A]);
+ pins.getState().endPinnedConnectedDrag(true);
+ assert.deepEqual(pins.getState().pinned, [C, B, A]);
+ assert.deepEqual(storedOrder(), [C, B, A]);
+});
+
+test("storage events outside a drag apply immediately", () => {
+ reset();
+ externalWrite([C, A]);
+ assert.deepEqual(pins.getState().pinned, [C, A]);
+});
diff --git a/tests/studio/studiobench/scene/actions.py b/tests/studio/studiobench/scene/actions.py
index c69cbfe8003..cd1d8e5fd3c 100644
--- a/tests/studio/studiobench/scene/actions.py
+++ b/tests/studio/studiobench/scene/actions.py
@@ -1075,14 +1075,7 @@ def settings(ctx: ActionContext) -> ActionResult:
@register_action(name = "model_change", default_budget_ms = 10000)
def model_change(ctx: ActionContext) -> ActionResult:
- """Open the model picker and select a row.
-
- THE WEAKEST SELECTOR IN THE SUITE, and it is recorded as such rather than hidden. The picker's
- option rows are plain `<button>` elements with utility classes: no `role="option"`, no
- `data-model-id`, no CommandItem, nothing stable anywhere in features/model-picker. So the row
- is found by position among the menu's buttons and the assertion is on the trigger's LABEL
- changing, which is an observable consequence rather than a selector.
- """
+ """Open the model picker and select a row marked with data-model-picker-option."""
trigger = ctx.page.query_selector("button.unsloth-model-selector-trigger")
if trigger is None:
return not_run("no model selector trigger on the page")
@@ -1129,7 +1122,7 @@ def model_change(ctx: ActionContext) -> ActionResult:
"label_before": before,
"label_after": after,
"menu_closed": closed,
- "selector_confidence": "low: the option rows carry no stable attribute",
+ "selector_confidence": "high: data-model-picker-option identifies selectable rows",
},
timings = {"open_ms": round(opened_ms, 1), "select_ms": round(select_ms, 1)},
reason = None if ok else "no option could be selected, or the menu stayed open",
diff --git a/tests/studio/studiobench/scene/dom.js b/tests/studio/studiobench/scene/dom.js
index 3f41be6aa5a..1a092451a38 100644
--- a/tests/studio/studiobench/scene/dom.js
+++ b/tests/studio/studiobench/scene/dom.js
@@ -331,9 +331,7 @@
},
modelOptions() {
const menu = D.modelMenu();
- // No role="option", no data-model-id: the rows are plain buttons with utility classes, so
- // this is the only available handle and it is recorded as the weak point it is.
- return menu ? qa("button", menu) : [];
+ return menu ? qa("button[data-model-picker-option]", menu) : [];
},
currentModelLabel() {
const t = D.modelTrigger();
diff --git a/tests/studio/studiobench/scene/selftest/test_studiobench_model_options.py b/tests/studio/studiobench/scene/selftest/test_studiobench_model_options.py
new file mode 100644
index 00000000000..97bfa01c9d0
--- /dev/null
+++ b/tests/studio/studiobench/scene/selftest/test_studiobench_model_options.py
@@ -0,0 +1,54 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Model selection excludes connection settings and row action buttons."""
+
+import shutil
+import subprocess
+from pathlib import Path
+
+import pytest
+
+
+def test_model_options_only_returns_selectable_rows():
+ node = shutil.which("node")
+ if node is None:
+ pytest.skip("node is not installed")
+ dom_js = Path(__file__).resolve().parents[1] / "dom.js"
+ script = r"""
+const assert = require("node:assert/strict");
+const fs = require("node:fs");
+const vm = require("node:vm");
+const buttons = [
+ { label: "Connections", option: false },
+ { label: "Model A", option: true },
+ { label: "Model settings", option: false },
+ { label: "Model B", option: true },
+];
+let menuOpen = true;
+const menu = {
+ querySelectorAll(selector) {
+ if (selector === "button") return buttons;
+ if (selector === "button[data-model-picker-option]") {
+ return buttons.filter((button) => button.option);
+ }
+ throw new Error(`Unexpected selector: ${selector}`);
+ },
+};
+const context = {
+ setInterval() {},
+ window: { addEventListener() {} },
+ document: {
+ querySelector(selector) {
+ assert.equal(selector, ".unsloth-model-selector-menu");
+ return menuOpen ? menu : null;
+ },
+ },
+};
+vm.runInNewContext(fs.readFileSync(process.argv[1], "utf8"), context);
+const options = context.window.__sb.dom.modelOptions();
+assert.deepEqual(Array.from(options, (button) => button.label), ["Model A", "Model B"]);
+menuOpen = false;
+assert.equal(context.window.__sb.dom.modelOptions().length, 0);
+"""
+ subprocess.run([node, "-e", script, str(dom_js)], check = True, timeout = 10)