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: model picker hides empty uploads and fades models this device cannot run (#14935)

Michael Han9f59a73ff908
diff --git a/studio/frontend/src/features/hub/hooks/use-hub-model-search.ts b/studio/frontend/src/features/hub/hooks/use-hub-model-search.ts
index 65bfc7eaa10..e59c6c9b1c9 100644
--- a/studio/frontend/src/features/hub/hooks/use-hub-model-search.ts
+++ b/studio/frontend/src/features/hub/hooks/use-hub-model-search.ts
@@ -21,6 +21,7 @@ import { EMBEDDING_TAGS, estimateSizeFromDtypes } from "../lib/hf-model-meta";
import { mergeTaskIterators } from "../lib/merge-task-iterators";
import { detectBaseModel } from "../lib/model-capabilities";
import { isGgufLike } from "../lib/model-identifiers";
+import { ggufRepoParams } from "@/lib/model-size";
import { type FreshRankable, rankFresh } from "../lib/fresh-ranking";
import { type ProviderLogo, resolveOwnerProviderLogo } from "../lib/provider-logos";
import {
@@ -190,6 +191,7 @@ function makeMapModel(
keepUnsupportedTags: boolean,
idSuffix: string,
deviceType: string | null,
+ hideEmptyUploads = false,
) {
const suffixLower = idSuffix.toLowerCase();
return (raw: unknown): HfModelResult | null => {
@@ -228,6 +230,15 @@ function makeMapModel(
if (excludeGguf && isGguf) {
return null;
}
+ // The Hub derives gguf metadata and the "gguf" tag from uploaded files: neither means none yet.
+ if (
+ hideEmptyUploads &&
+ isGgufLike(m.name) &&
+ !m.gguf &&
+ !m.tags?.some((tag) => tag.toLowerCase() === "gguf")
+ ) {
+ return null;
+ }
const pipelineTag = m.task ?? m.pipeline_tag;
const quantMethod = m.config?.quantization_config?.quant_method;
// Drop runtime-unloadable models before they reach the row list. Discover opts out via
@@ -261,7 +272,9 @@ function makeMapModel(
likes: m.likes ?? 0,
private: m.private,
gated: m.gated,
- totalParams: m.safetensors?.total ?? m.gguf?.total,
+ totalParams:
+ m.safetensors?.total ??
+ (m.gguf ? ggufRepoParams(m.name, m.gguf.total) : undefined),
estimatedSizeBytes: estimateSizeFromDtypes(m.safetensors?.parameters),
isGguf,
baseModel: detectBaseModel(m.tags),
@@ -859,6 +872,8 @@ export function useHubModelSearch(
ownerScope?: "unsloth" | "all";
enabled?: boolean;
keepUnsupportedTags?: boolean;
+ /** Drop "-GGUF" repos with no GGUF files uploaded yet. The Hub page keeps them listed. */
+ hideEmptyUploads?: boolean;
channel?: HfModelSearchChannel | null;
filters?: ModelSearchFilters;
},
@@ -874,6 +889,7 @@ export function useHubModelSearch(
ownerScope = "all",
enabled = true,
keepUnsupportedTags = false,
+ hideEmptyUploads = false,
channel = null,
filters,
} = options ?? {};
@@ -1172,12 +1188,14 @@ export function useHubModelSearch(
keepUnsupportedTags,
channelIdSuffix,
deviceType,
+ hideEmptyUploads,
),
[
excludeGguf,
keepUnsupportedTags,
channelIdSuffix,
deviceType,
+ hideEmptyUploads,
],
);
const search = useHubPaginatedSearch(createIter, mapModel, { enabled });
diff --git a/studio/frontend/src/features/hub/lib/model-search-filters.ts b/studio/frontend/src/features/hub/lib/model-search-filters.ts
index ed1f89e0b53..adab7ae11dc 100644
--- a/studio/frontend/src/features/hub/lib/model-search-filters.ts
+++ b/studio/frontend/src/features/hub/lib/model-search-filters.ts
@@ -1,6 +1,8 @@
// 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 { ggufRepoParams } from "../../../lib/model-size.ts";
+
/** Who uploaded a model: the unsloth org, any other organization, or a personal account. */
export type HubUploader = "unsloth" | "organization" | "personal";
@@ -115,7 +117,10 @@ export async function* filterModelListing(
const model = raw as ListingModel;
if (
!matchesRange(
- model.safetensors?.total ?? model.gguf?.total,
+ model.safetensors?.total ??
+ (model.gguf
+ ? ggufRepoParams(model.name, model.gguf.total)
+ : undefined),
filters.minParams,
filters.maxParams,
)
diff --git a/studio/frontend/src/features/hub/lib/view-models.ts b/studio/frontend/src/features/hub/lib/view-models.ts
index 00ed84cfbe2..ec7a4dfd554 100644
--- a/studio/frontend/src/features/hub/lib/view-models.ts
+++ b/studio/frontend/src/features/hub/lib/view-models.ts
@@ -14,6 +14,7 @@ import type {
ModelFormatFilter,
} from "../types";
import { estimateSizeFromDtypes, isGgufLike } from "./hf-model-meta";
+import { ggufRepoParams } from "@/lib/model-size";
import {
type CapabilityKey,
detectBaseModel,
@@ -187,7 +188,10 @@ export function toHfModelResult(raw: unknown): HfModelResult | null {
private: model.private,
gated: model.gated,
totalParams:
- finiteNumber(model.safetensors?.total) ?? finiteNumber(model.gguf?.total),
+ finiteNumber(model.safetensors?.total) ??
+ (model.gguf
+ ? ggufRepoParams(model.name, finiteNumber(model.gguf.total))
+ : undefined),
estimatedSizeBytes: estimateSizeFromDtypes(model.safetensors?.parameters),
isGguf,
baseModel: detectBaseModel(model.tags),
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 b4383211792..5390828574c 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
@@ -113,6 +113,7 @@ import {
type ModelMemorySource,
useModelMemory,
} from "@/hooks/use-model-memory";
+import type { GpuInfo } from "@/hooks/use-gpu-info";
import { useVramBudgetFraction } from "@/hooks/use-vram-budget-fraction";
import { copyToClipboard } from "@/lib/copy-to-clipboard";
import { diffusionRouteSearch } from "@/lib/diffusion-route-search";
@@ -870,6 +871,19 @@ function isOverBudget(status?: GgufFitClass | VramFitStatus | null): boolean {
);
}
+/** A checkpoint (non-GGUF) row's verdict. A CPU backend chats with GGUF only, so a checkpoint
+ * cannot run there at any size; elsewhere no GPU means nothing to judge it against. */
+function checkpointFit(
+ estGb: number,
+ gpu: Pick<GpuInfo, "available" | "backend" | "budgetKnown" | "memoryTotalGb">,
+ chat: boolean,
+): VramFitStatus | null {
+ if (chat && gpu.budgetKnown && gpu.backend === "cpu") return "exceeds";
+ return estGb > 0 && gpu.available
+ ? checkVramFit(estGb, gpu.memoryTotalGb)
+ : null;
+}
+
/** VRAM verdict: an info mark that names itself on hover, rather than a shouted pill. */
function VramBadge({
status,
@@ -1274,6 +1288,8 @@ function ModelRow({
className?: string;
}) {
const exceeds = isOverBudget(vramStatus);
+ // With no GPU every load runs from RAM: `ram` keeps its warning but is not an overage to dim.
+ const dimmed = exceeds && !(vramStatus === "ram" && !(gpuGb && gpuGb > 0));
const showVramTooltip =
vramEst != null && vramEst > 0 && gpuGb != null && gpuGb > 0;
const vramTooltipText =
@@ -1434,9 +1450,9 @@ function ModelRow({
"flex w-full items-center",
alignMeta === "device" ? DEVICE_META_GAP : "gap-1",
// Over budget reads as a dimmed row, which scans; hover restores it. The selected row keeps full weight.
- exceeds &&
+ dimmed &&
!selected &&
- "opacity-60 transition-opacity group-hover/row:opacity-100 group-focus-visible/row:opacity-100",
+ "opacity-55 transition-opacity group-hover/row:opacity-100 group-focus-visible/row:opacity-100 dark:opacity-50",
)}
>
<span className="flex min-w-0 flex-1 items-baseline">
@@ -3289,6 +3305,7 @@ export function HubModelPicker({
sortDirection: "desc",
pinUnslothFirst: true,
keepUnsupportedTags: true,
+ hideEmptyUploads: true,
accessToken,
// Only Recommended renders Hub results, so keep the Hub hooks idle on other tabs and preserve
// offline-local behavior. Search rows render only with a query (showHfSection); browse reads
@@ -3303,6 +3320,7 @@ export function HubModelPicker({
sortDirection: "desc",
pinUnslothFirst: true,
keepUnsupportedTags: true,
+ hideEmptyUploads: true,
accessToken,
enabled: online && section === "recommended",
});
@@ -3324,6 +3342,7 @@ export function HubModelPicker({
sortBy: recommendedSort,
sortDirection: "desc",
pinUnslothFirst: false,
+ hideEmptyUploads: true,
accessToken,
enabled: communityDiscoveryEnabled && debouncedQuery.trim().length > 0,
});
@@ -3333,6 +3352,7 @@ export function HubModelPicker({
sortBy: recommendedSort,
sortDirection: "desc",
pinUnslothFirst: false,
+ hideEmptyUploads: true,
accessToken,
enabled: communityRecommendedEnabled && debouncedQuery.trim().length === 0,
});
@@ -4368,9 +4388,7 @@ export function HubModelPicker({
const est = r.totalParams
? estimateLoadingVram(r.totalParams, "qlora")
: 0;
- const status =
- est > 0 && gpu.available ? checkVramFit(est, gpu.memoryTotalGb) : null;
- map.set(r.id, { meta, status, est });
+ map.set(r.id, { meta, status: checkpointFit(est, gpu, !task), est });
}
return map;
}, [
@@ -5998,16 +6016,13 @@ export function HubModelPicker({
const detail = r.totalParams ? formatCompact(r.totalParams) : null;
if (r.totalParams) {
const est = estimateLoadingVram(r.totalParams, "qlora");
- const status = gpu.available
- ? checkVramFit(est, gpu.memoryTotalGb)
- : null;
- map.set(r.id, { est, status, detail });
+ map.set(r.id, { est, status: checkpointFit(est, gpu, !task), detail });
} else {
map.set(r.id, { est: 0, status: null, detail });
}
}
return map;
- }, [results, gpu]);
+ }, [results, gpu, task]);
const recommendedVramMap = useMemo(() => {
const map = new Map<
@@ -6041,11 +6056,8 @@ export function HubModelPicker({
}
if (totalParams) {
const est = estimateLoadingVram(totalParams, "qlora");
- const status = gpu.available
- ? checkVramFit(est, gpu.memoryTotalGb)
- : null;
const detail = formatCompact(totalParams);
- map.set(id, { est, status, detail });
+ map.set(id, { est, status: checkpointFit(est, gpu, !task), detail });
}
}
return map;
diff --git a/studio/frontend/src/lib/model-size.ts b/studio/frontend/src/lib/model-size.ts
index 5e0e5826398..a335674aef0 100644
--- a/studio/frontend/src/lib/model-size.ts
+++ b/studio/frontend/src/lib/model-size.ts
@@ -21,3 +21,15 @@ export function parseParamCountB(id: string): number | null {
const v = Number.parseFloat(m[1]);
return Number.isFinite(v) ? v : null;
}
+
+/** A GGUF repo's parameter count. The Hub reads `gguf.total` from one file of its choosing, which
+ * can be a drafter or vision projector (1.6B on a 120B repo), so a larger name token wins. */
+export function ggufRepoParams(
+ id: string,
+ ggufTotal: number | undefined,
+): number | undefined {
+ const fromName = (parseParamCountB(id) ?? 0) * 1e9;
+ const total = ggufTotal ?? 0;
+ const largest = Math.max(fromName, total);
+ return largest > 0 ? largest : undefined;
+}
diff --git a/studio/frontend/tests/hub-search-empty-uploads.test.ts b/studio/frontend/tests/hub-search-empty-uploads.test.ts
new file mode 100644
index 00000000000..4d0f39abae1
--- /dev/null
+++ b/studio/frontend/tests/hub-search-empty-uploads.test.ts
@@ -0,0 +1,82 @@
+// 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 { readFileSync } from "node:fs";
+import test from "node:test";
+import ts from "@typescript/typescript6";
+import { classifyUnslothSupport } from "../src/features/hub/lib/unsloth-support.ts";
+import { ggufRepoParams } from "../src/lib/model-size.ts";
+
+const source = readFileSync(
+ new URL("../src/features/hub/hooks/use-hub-model-search.ts", import.meta.url),
+ "utf8",
+);
+const start = source.indexOf("function makeMapModel(");
+assert.notEqual(start, -1);
+const end = source.indexOf("\n}\n", start) + 2;
+const makeMapModel = new Function(
+ "classifyUnslothSupport", "EMBEDDING_TAGS", "isGgufLike", "estimateSizeFromDtypes", "detectBaseModel", "ggufRepoParams",
+ ts.transpileModule(source.slice(start, end), { compilerOptions: { target: ts.ScriptTarget.ES2022 } }).outputText +
+ "; return makeMapModel;",
+)(
+ classifyUnslothSupport,
+ new Set<string>(),
+ (id: string) => /-gguf$/i.test(id),
+ () => undefined,
+ () => null,
+ ggufRepoParams,
+);
+
+// unsloth/MiniMax-M3.1-Flash-GGUF as the listing reported it while only README.md was uploaded.
+const PLACEHOLDER = {
+ name: "unsloth/MiniMax-M3.1-Flash-GGUF",
+ pipeline_tag: "image-text-to-text",
+ tags: ["unsloth", "image-text-to-text", "base_model:MiniMaxAI/MiniMax-M3.1-Flash", "region:us"],
+};
+const UPLOADED = [
+ { name: "unsloth/flint-GGUF", gguf: { total: 39785287456 }, tags: ["gguf", "unsloth"] },
+ // A mirror that drops expand=gguf still reports the tag.
+ { name: "unsloth/Qwen3-8B-GGUF", tags: ["gguf"] },
+ // A non-GGUF repo is never judged by this rule.
+ { name: "unsloth/FLUX.1-dev-FP8", tags: ["diffusers"] },
+ { name: "unsloth/MiniMax-M3.1-Flash", safetensors: { total: 220490854016 }, tags: [] },
+];
+
+test("the picker drops a GGUF repo with nothing uploaded yet", () => {
+ const map = makeMapModel(false, true, "", "cuda", true);
+ assert.equal(map(PLACEHOLDER), null);
+ for (const raw of UPLOADED) assert.equal(map(raw)?.id, raw.name);
+});
+
+test("listings that do not opt in keep the placeholder", () => {
+ assert.equal(makeMapModel(false, true, "", "cuda")(PLACEHOLDER)?.id, PLACEHOLDER.name);
+});
+
+test("every Hub listing in the model picker opts in", () => {
+ const picker = readFileSync(
+ new URL("../src/features/model-picker/components/model-selector/pickers.tsx", import.meta.url),
+ "utf8",
+ );
+ const calls = picker.split("useHubModelSearch(").slice(1);
+ assert.ok(calls.length >= 4);
+ for (const call of calls) {
+ const options = call.slice(0, call.indexOf("});"));
+ assert.match(options, /hideEmptyUploads: true/);
+ }
+});
+
+test("a GGUF repo's size is its largest reading, not the drafter the Hub sampled", () => {
+ const map = makeMapModel(false, true, "", "cuda", true);
+ // The Hub's gguf.total here comes from a 1.6B DFlash drafter shipped beside the 120B quants.
+ const nemotron = map({
+ name: "unsloth/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF",
+ gguf: { total: 1632922404, architecture: "dflash" },
+ tags: ["gguf"],
+ });
+ assert.equal(nemotron?.totalParams, 120e9);
+ // Metadata stays when it is the larger, more exact figure, or the name has no size.
+ assert.equal(map({ name: "unsloth/Qwen3-30B-A3B-GGUF", gguf: { total: 30.5e9 }, tags: ["gguf"] })?.totalParams, 30.5e9);
+ assert.equal(map({ name: "unsloth/flint-GGUF", gguf: { total: 39785287456 }, tags: ["gguf"] })?.totalParams, 39785287456);
+ // Safetensors totals already count every weight and keep precedence.
+ assert.equal(map({ name: "unsloth/MiniMax-M3.1-Flash", safetensors: { total: 220490854016 }, tags: [] })?.totalParams, 220490854016);
+});
diff --git a/studio/frontend/tests/hub-search-mixed-formats.test.ts b/studio/frontend/tests/hub-search-mixed-formats.test.ts
index c3ce545c8b9..fe978359c61 100644
--- a/studio/frontend/tests/hub-search-mixed-formats.test.ts
+++ b/studio/frontend/tests/hub-search-mixed-formats.test.ts
@@ -5,6 +5,7 @@ import { readFileSync } from "node:fs";
import test from "node:test";
import ts from "@typescript/typescript6";
import { classifyUnslothSupport } from "../src/features/hub/lib/unsloth-support.ts";
+import { ggufRepoParams } from "../src/lib/model-size.ts";
// Execute the production mapper without loading React, the Hub client or the network hooks.
const source = readFileSync(
@@ -15,7 +16,7 @@ const start = source.indexOf("function makeMapModel(");
assert.notEqual(start, -1);
const end = source.indexOf("\n}\n", start) + 2;
const makeMapModel = new Function(
- "classifyUnslothSupport", "EMBEDDING_TAGS", "isGgufLike", "estimateSizeFromDtypes", "detectBaseModel",
+ "classifyUnslothSupport", "EMBEDDING_TAGS", "isGgufLike", "estimateSizeFromDtypes", "detectBaseModel", "ggufRepoParams",
ts.transpileModule(source.slice(start, end), { compilerOptions: { target: ts.ScriptTarget.ES2022 } }).outputText +
"; return makeMapModel;",
)(
@@ -24,6 +25,7 @@ const makeMapModel = new Function(
(id: string) => id.endsWith("-GGUF"),
() => undefined,
() => null,
+ ggufRepoParams,
);
const GPT2 = ["pytorch", "tf", "jax", "tflite", "onnx", "safetensors"];
diff --git a/studio/frontend/tests/on-device-dot-and-header-alignment.test.ts b/studio/frontend/tests/on-device-dot-and-header-alignment.test.ts
index 2bed0dcac9b..c23d3085ef9 100644
--- a/studio/frontend/tests/on-device-dot-and-header-alignment.test.ts
+++ b/studio/frontend/tests/on-device-dot-and-header-alignment.test.ts
@@ -167,7 +167,7 @@ test("an over budget row dims instead of putting a pill on every line", () => {
);
assert.ok(
PICKERS.includes(
- '"opacity-60 transition-opacity group-hover/row:opacity-100 group-focus-visible/row:opacity-100"',
+ '"opacity-55 transition-opacity group-hover/row:opacity-100 group-focus-visible/row:opacity-100 dark:opacity-50"',
),
);
assert.ok(PICKERS.includes('vram: "min-w-min min-[560px]:w-[18px]"'));
@@ -191,7 +191,7 @@ test("one fit badge, so a colour or reveal change cannot miss a list", () => {
2,
"both model row slots reveal on hover",
);
- assert.match(PICKERS, /exceeds &&\n\s*!selected &&/);
+ assert.match(PICKERS, /dimmed &&\n\s*!selected &&/);
});
test("chat and the Hub answer the fit question with one formula", () => {
diff --git a/studio/frontend/tests/picker-cpu-fit.test.ts b/studio/frontend/tests/picker-cpu-fit.test.ts
new file mode 100644
index 00000000000..e2a3aa08f32
--- /dev/null
+++ b/studio/frontend/tests/picker-cpu-fit.test.ts
@@ -0,0 +1,48 @@
+// 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 { readFileSync } from "node:fs";
+import test from "node:test";
+import ts from "@typescript/typescript6";
+import { checkVramFit } from "../src/lib/vram.ts";
+
+const PICKERS = readFileSync(
+ new URL("../src/features/model-picker/components/model-selector/pickers.tsx", import.meta.url),
+ "utf8",
+);
+const start = PICKERS.indexOf("function checkpointFit(");
+assert.notEqual(start, -1);
+const end = PICKERS.indexOf("\n}\n", start) + 2;
+const checkpointFit = new Function(
+ "checkVramFit",
+ ts.transpileModule(PICKERS.slice(start, end), { compilerOptions: { target: ts.ScriptTarget.ES2022 } }).outputText +
+ "; return checkpointFit;",
+)(checkVramFit);
+
+// What /api/system reports on a GGUF-only Mac (no torch, so no MLX device).
+const CPU = { available: false, backend: "cpu", budgetKnown: true, memoryTotalGb: 0 };
+
+test("a checkpoint cannot run on a CPU backend, so its row dims", () => {
+ assert.equal(checkpointFit(130, CPU, true), "exceeds");
+ assert.equal(checkpointFit(0, CPU, true), "exceeds");
+});
+
+test("a checkpoint keeps its old verdicts everywhere else", () => {
+ // Still loading the inventory: no verdict yet.
+ assert.equal(checkpointFit(130, { ...CPU, budgetKnown: false }, true), null);
+ // Task pickers keep their own rules.
+ assert.equal(checkpointFit(130, CPU, false), null);
+ const cuda = { available: true, backend: "cuda", budgetKnown: true, memoryTotalGb: 80 };
+ assert.equal(checkpointFit(130, cuda, true), "exceeds");
+ assert.equal(checkpointFit(20, cuda, true), "fits");
+ assert.equal(checkpointFit(130, { ...cuda, available: false, backend: "mlx" }, true), null);
+});
+
+test("with no GPU a GGUF that fits in RAM keeps its warning but is not dimmed", () => {
+ // The verdict still reaches the row, so the RAM fallback mark shows.
+ assert.doesNotMatch(PICKERS, /fit === "ram" && budget\.memoryTotalGb <= 0/);
+ assert.match(
+ PICKERS,
+ /const dimmed = exceeds && !\(vramStatus === "ram" && !\(gpuGb && gpuGb > 0\)\);/,
+ );
+});