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: show the newest Kimi models in the model picker (#16591)

Nilay37e942958e32
diff --git a/studio/backend/core/inference/external_provider.py b/studio/backend/core/inference/external_provider.py
index 3e50a0ed039..f26c4f38d2b 100644
--- a/studio/backend/core/inference/external_provider.py
+++ b/studio/backend/core/inference/external_provider.py
@@ -903,10 +903,11 @@ def _apply_deepseek_reasoning_controls(
_KIMI_K3_EFFORT_ALIASES = {"minimal": "low", "medium": "high", "xhigh": "max"}
+_KIMI_K3_FAMILY = re.compile(r"kimi-k3(?:$|[.-])")
def _is_kimi_k3(provider_type: Optional[str], model: Optional[str]) -> bool:
- return provider_type == "kimi" and (model or "").strip().lower() == "kimi-k3"
+ return provider_type == "kimi" and bool(_KIMI_K3_FAMILY.match((model or "").strip().lower()))
def _apply_kimi_k3_controls(body: dict[str, Any], reasoning_effort: Optional[str]) -> None:
diff --git a/studio/backend/core/inference/providers.py b/studio/backend/core/inference/providers.py
index dc592bfe150..8dd75230cd8 100644
--- a/studio/backend/core/inference/providers.py
+++ b/studio/backend/core/inference/providers.py
@@ -217,13 +217,12 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
"kimi": {
"display_name": "Kimi",
"base_url": "https://api.moonshot.ai/v1",
- # Surface only the SoTA multimodal models (kimi-k3, kimi-k2.6/k2.5);
- # moonshot-v1-* and dated k2 previews are filtered by the allowlist.
# Docs: https://platform.kimi.ai/docs/models
# Listing/overview: https://platform.kimi.ai/docs/api/list-models
# https://platform.kimi.ai/docs/api/overview
"default_models": [
"kimi-k3",
+ "kimi-k3.1-pro",
"kimi-k2.6",
"kimi-k2.5",
],
@@ -236,7 +235,8 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
"auth_header": "Authorization",
"auth_prefix": "Bearer ",
"notes": "Moonshot API key. China: use base URL https://api.moonshot.cn/v1",
- "model_id_allowlist": re.compile(r"^kimi-(?:k3|k2\.[56])$"),
+ "model_id_allowlist": re.compile(r"^kimi-k\d"),
+ "model_id_denylist": re.compile(r"-preview$|^kimi-k2-thinking"),
# Reasoning-class: the API rejects custom temperature/top_p ("only 1
# is allowed"). Strip both so the server uses its required defaults.
"body_omit": ("temperature", "top_p"),
diff --git a/studio/backend/tests/test_kimi_k3_provider.py b/studio/backend/tests/test_kimi_k3_provider.py
index ffd48ce2e95..9ec305acafe 100644
--- a/studio/backend/tests/test_kimi_k3_provider.py
+++ b/studio/backend/tests/test_kimi_k3_provider.py
@@ -57,23 +57,69 @@ def _stream(model, **kwargs):
_drive(run())
+def _list_kimi_models(monkeypatch, ids):
+ from models.providers import ProviderModelsRequest
+ from routes import providers as providers_route
+
+ class _FakeClient:
+ def __init__(self, **_kwargs):
+ pass
+
+ async def list_models(self):
+ return [{"id": model_id} for model_id in ids]
+
+ async def close(self):
+ return None
+
+ monkeypatch.setattr(providers_route, "ExternalProviderClient", _FakeClient)
+ monkeypatch.setattr(
+ providers_route, "resolve_provider_api_key_or_400", lambda *a, **k: "sk-test"
+ )
+ payload = ProviderModelsRequest(provider_type = "kimi", base_url = "https://api.moonshot.ai/v1")
+ result = _drive(providers_route.list_provider_models(payload, "tester", False))
+ return [m.id for m in result]
+
+
+_MOONSHOT_LIVE_IDS = [
+ "kimi-k3",
+ "kimi-k2.7-code",
+ "kimi-k2.7-code-highspeed",
+ "kimi-k2.6",
+ "kimi-k3.1-pro",
+]
+
+
def test_kimi_k3_reaches_the_picker():
kimi = PROVIDER_REGISTRY["kimi"]
- for model_id in ("kimi-k3", "kimi-k2.6", "kimi-k2.5"):
- assert kimi["model_id_allowlist"].match(model_id), model_id
+ for model_id in ("kimi-k3", "kimi-k3.1-pro", "kimi-k2.6", "kimi-k2.5"):
assert model_id in kimi["default_models"], model_id
- for model_id in ("kimi-k3-preview", "kimi-k2-0905-preview", "moonshot-v1-8k"):
- assert not kimi["model_id_allowlist"].match(model_id), model_id
+
+
+def test_every_current_kimi_model_reaches_the_picker(monkeypatch):
+ assert _list_kimi_models(monkeypatch, _MOONSHOT_LIVE_IDS) == _MOONSHOT_LIVE_IDS
+
+
+def test_kimi_previews_and_legacy_models_stay_out_of_the_picker(monkeypatch):
+ ids = [
+ "kimi-k3-preview",
+ "kimi-k2-0905-preview",
+ "kimi-k2-turbo-preview",
+ "kimi-k2-thinking",
+ "kimi-k2-thinking-turbo",
+ "moonshot-v1-8k",
+ ]
+ assert _list_kimi_models(monkeypatch, ids) == []
@pytest.mark.parametrize(
("effort", "sent"),
[("low", "low"), ("high", "high"), ("max", "max"), ("medium", "high"), ("xhigh", "max")],
)
-def test_kimi_k3_sends_its_documented_controls(monkeypatch, effort, sent):
[email protected]("model", ["kimi-k3", "kimi-k3.1-pro"])
+def test_kimi_k3_sends_its_documented_controls(monkeypatch, model, effort, sent):
bodies = _capture(monkeypatch, [_PLAIN])
_stream(
- "kimi-k3",
+ model,
temperature = 0.7,
top_p = 0.9,
presence_penalty = 0.0,
@@ -161,7 +207,9 @@ def test_kimi_k3_history_keeps_prior_reasoning():
assert all("reasoning_content" not in m for m in k26)
[email protected](("model", "keeps"), [("kimi-k3", True), ("kimi-k2.6", False)])
+ ("model", "keeps"), [("kimi-k3", True), ("kimi-k3.1-pro", True), ("kimi-k2.6", False)]
+)
def test_studio_tool_loop_replays_kimi_k3_reasoning(model, keeps):
from core.inference.external_tool_transport import OAICompatTransport
client = ExternalProviderClient(
diff --git a/studio/frontend/src/features/chat/provider-capabilities.ts b/studio/frontend/src/features/chat/provider-capabilities.ts
index 49d805c256f..24fbc0e59bd 100644
--- a/studio/frontend/src/features/chat/provider-capabilities.ts
+++ b/studio/frontend/src/features/chat/provider-capabilities.ts
@@ -352,14 +352,16 @@ export function providerSupportsBuiltinWebSearch(
);
}
+function isKimiK3(modelId: string | null | undefined): boolean {
+ return /^kimi-k3(?:$|[.-])/.test(modelId?.trim().toLowerCase() ?? "");
+}
+
/** Search and Think stay exclusive on Kimi, except kimi-k3, which has no thinking off switch. */
export function kimiSearchExcludesThinking(
providerType: string | null | undefined,
modelId: string | null | undefined,
): boolean {
- return (
- providerType === "kimi" && modelId?.trim().toLowerCase() !== "kimi-k3"
- );
+ return providerType === "kimi" && !isKimiK3(modelId);
}
/** Whether the provider exposes a server-side web_fetch tool emitting a document block.
@@ -722,7 +724,7 @@ export function getProviderCapabilities(
};
}
// kimi-k3 fixes presence_penalty too and the docs ask for it to be omitted.
- if (providerType === "kimi" && modelId?.trim().toLowerCase() === "kimi-k3") {
+ if (providerType === "kimi" && isKimiK3(modelId)) {
return { ...PROVIDER_CAPABILITIES.kimi, presencePenalty: false };
}
return PROVIDER_CAPABILITIES[providerType] ?? DEFAULT_EXTERNAL_CAPABILITIES;
@@ -952,7 +954,7 @@ function resolveKimiReasoningCapabilities(modelId: string): ExternalReasoningCap
// toggleable, k2-thinking always on, k2.5 and others none.
// k2.6 is toggled via extra_body: {thinking: {type: enabled|disabled}}. k3 always reasons
// and takes reasoning_effort low|high|max instead.
- if (modelId === "kimi-k3") {
+ if (isKimiK3(modelId)) {
return {
...withReasoningEffortStyle({
supportsReasoning: true,
@@ -1380,7 +1382,7 @@ export function providerModelReplaysReasoning(
): boolean {
return (
providerType === "minimax" ||
- (providerType === "kimi" && modelId?.trim().toLowerCase() === "kimi-k3")
+ (providerType === "kimi" && isKimiK3(modelId))
);
}
diff --git a/studio/frontend/tests/provider-capabilities-current-models.test.ts b/studio/frontend/tests/provider-capabilities-current-models.test.ts
index ee461a35c14..cf7d9daf452 100644
--- a/studio/frontend/tests/provider-capabilities-current-models.test.ts
+++ b/studio/frontend/tests/provider-capabilities-current-models.test.ts
@@ -566,6 +566,11 @@ test("Kimi K3 always thinks at a low/high/max effort and leaves k2.x as they wer
);
assert.equal(providerModelReplaysReasoning("kimi", "kimi-k3"), true);
+ assert.deepEqual(getExternalReasoningCapabilities("kimi", "kimi-k3.1-pro"), k3);
+ assert.equal(getProviderCapabilities("kimi", undefined, "kimi-k3.1-pro")?.presencePenalty, false);
+ assert.equal(kimiSearchExcludesThinking("kimi", "kimi-k3.1-pro"), false);
+ assert.equal(providerModelReplaysReasoning("kimi", "kimi-k3.1-pro"), true);
+
const k26 = getExternalReasoningCapabilities("kimi", "kimi-k2.6");
assert.equal(k26.reasoningStyle, "enable_thinking");
assert.equal(k26.supportsReasoningOff, true);