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: pin MiniMax-H3-Image (Fast FP8) at the top of the image picker (#17502)

Daniel Han475ceee98d8d
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/model-catalog.check.ts b/studio/frontend/src/features/model-picker/components/model-selector/model-catalog.check.ts
index 3453aed255b..2b2656d8c8d 100644
--- a/studio/frontend/src/features/model-picker/components/model-selector/model-catalog.check.ts
+++ b/studio/frontend/src/features/model-picker/components/model-selector/model-catalog.check.ts
@@ -160,10 +160,40 @@ for (const [base, fp8Repo, name] of [
`${base.split("/")[1]}-GGUF (Slow)`,
);
}
-assert.equal(
- IMAGE_CATALOG.find((g) => g.pinToTop)?.canonicalId,
- "black-forest-labs/flux-3-image-klein",
+// MiniMax-H3-Image leads the pinned order, FLUX.3 Image Klein right under it, then Qwen-Image 2.1 and its Turbo.
+assert.deepEqual(
+ IMAGE_CATALOG.filter((g) => g.pinToTop).map((g) => g.canonicalId),
+ ["MiniMaxAI/MiniMax-H3-Image", "black-forest-labs/flux-3-image-klein", "unsloth/Qwen-Image-2.1", "Qwen/Qwen-Image-2.1-Turbo"],
);
+{
+ const h3i = "MiniMaxAI/MiniMax-H3-Image";
+ const group = groupForRepoId(h3i, IMAGE_CATALOG);
+ assert.ok(group);
+ assert.equal(groupForRepoId("unsloth/MiniMax-H3-Image-FP8", IMAGE_CATALOG), group);
+ // The dense row IS the Fast FP8 row: no unpinned bf16 twin.
+ assert.equal(group.artifacts.length, 1);
+ assert.equal(group.artifacts[0].prequantRepo, "unsloth/MiniMax-H3-Image-FP8");
+ assert.deepEqual(group.artifacts[0].prequantSizeGb, { fp8: 18.88, int8: 18.88 });
+ assert.equal(group.artifacts[0].fetchedFrom, undefined);
+ for (const schemes of [["int8", "fp8"], ["fp8"], ["int8"]]) {
+ assert.equal(curatedRowLabelFor(h3i, IMAGE_CATALOG, "dense-quant", schemes)?.name, "MiniMax-H3-Image (Fast FP8)");
+ }
+ // The backend refuses Apple Silicon and CPU.
+ for (const host of ["accelerated", "gguf-only", "unknown"] as const) {
+ assert.equal(curatedRowLabelFor(h3i, IMAGE_CATALOG, host)?.name, "MiniMax-H3-Image", host);
+ }
+ const names = catalogToModelOptions(IMAGE_CATALOG, "dense-quant", ["int8", "fp8"])
+ .filter((o) => IMAGE_CATALOG.some((g) => g.pinToTop && groupForRepoId(o.id, IMAGE_CATALOG) === g))
+ .map((o) => o.name);
+ assert.deepEqual(names.slice(0, 2), ["MiniMax-H3-Image (Fast FP8)", "FLUX.3 Image Klein (Fast FP8)"]);
+ const fits = (gpuGb: number, systemRamGb: number, denseQuantSchemes?: string[]) =>
+ curatedArtifactFitsDevice(h3i, IMAGE_CATALOG, { gpuGb, systemRamGb, denseQuantSchemes });
+ assert.equal(fits(24, 64, ["int8", "fp8"]), true);
+ assert.equal(fits(40, 64, ["int8", "fp8"]), true);
+ assert.equal(fits(24, 64), false);
+ assert.equal(fits(24, 32, ["int8", "fp8"]), false);
+ assert.equal(fits(16, 64, ["int8", "fp8"]), false);
+}
// MOSS-Image: a dense row on the OpenMOSS id (fetched from the FP8 repo), with sd.cpp GGUF rows in the same group.
for (const [base, fp8Repo, ggufRepo, name] of [
["OpenMOSS-Team/MOSS-Image-SFT", "unsloth/MOSS-Image-SFT-FP8", "unsloth/MOSS-Image-SFT-GGUF", "MOSS-Image"],
@@ -1188,6 +1218,7 @@ const PREQUANT_ROWS = [
["black-forest-labs/flux-3-image-klein-base", "unsloth/flux-3-image-klein-base-FP8"],
["OpenMOSS-Team/MOSS-Image-SFT", "unsloth/MOSS-Image-SFT-FP8"],
["OpenMOSS-Team/MOSS-Image-Base", "unsloth/MOSS-Image-Base-FP8"],
+ ["MiniMaxAI/MiniMax-H3-Image", "unsloth/MiniMax-H3-Image-FP8"],
] as const;
for (const [id, repo] of PREQUANT_ROWS) {
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/model-catalog.ts b/studio/frontend/src/features/model-picker/components/model-selector/model-catalog.ts
index 0240ac0230b..889fc6ce248 100644
--- a/studio/frontend/src/features/model-picker/components/model-selector/model-catalog.ts
+++ b/studio/frontend/src/features/model-picker/components/model-selector/model-catalog.ts
@@ -217,6 +217,8 @@ const audioGgufGroups = (tasks: readonly AudioCppTask[]): CatalogGroup[] =>
const FLUX3_KLEIN_AUTO_TE = { denseGb: 8.27, quantisedGb: 4.89 } as const;
const QWEN_IMAGE_21_AUTO_TE = { denseGb: 16.33, quantisedGb: 8.71 } as const;
const FLUX1_KONTEXT_AUTO_TE = { denseGb: 8.87, quantisedGb: 5.74 } as const;
+// No te_quant_auto: the loader itself takes the hosted int8 Qwen3-VL-32B (unsloth/MiniMax-H3-FP8) when no encoder is set.
+const MINIMAX_H3_IMAGE_AUTO_TE = { denseGb: 62.14, quantisedGb: 25.28 } as const;
export const IMAGE_CATALOG: CatalogGroup[] = [
{
@@ -243,8 +245,32 @@ export const IMAGE_CATALOG: CatalogGroup[] = [
artifacts: [gguf("unsloth/Z-Image-GGUF")],
},
{
- // FLUX.3 Image Klein (distilled): text to image and up to 10 reference images. Pinned first, so its
- // "Fast FP8" and GGUF rows lead the list above Qwen-Image 2.1's. The bf16 row keeps the gated BFL id; the backend
+ // CUDA / ROCm only. Keeps the official id (loader reads VAE, tokenizer, config from it), so no fetchedFrom.
+ // Size is the bf16 total; the fit rule reads the measured offload tiers instead.
+ canonicalId: "MiniMaxAI/MiniMax-H3-Image",
+ displayName: "MiniMax-H3-Image",
+ description: "Text-to-image and image editing with up to 16 pictures",
+ scope: "image",
+ // Same reason as Qwen-Image 2.1's alias: the int8 half of the prequant repo has no artifact row.
+ aliases: ["unsloth/MiniMax-H3-Image-FP8"],
+ pinToTop: true,
+ pinRowsTogether: true,
+ artifacts: [
+ bf16Pipeline("MiniMaxAI/MiniMax-H3-Image", 126, {
+ totalParams: 33122643440,
+ prequantRepo: "unsloth/MiniMax-H3-Image-FP8",
+ prequantSizeGb: { fp8: 18.88, int8: 18.88 },
+ autoQuantTextEncoderGb: MINIMAX_H3_IMAGE_AUTO_TE,
+ offloadFitTiers: [
+ { gpuGb: 22, systemRamGb: 48, requiresDenseQuant: true },
+ { gpuGb: 39, systemRamGb: 64, requiresDenseQuant: true },
+ ],
+ }),
+ ],
+ },
+ {
+ // FLUX.3 Image Klein (distilled): text to image and up to 10 reference images. Pinned right under
+ // MiniMax-H3-Image. The bf16 row keeps the gated BFL id; the backend
// fetches it from unsloth/flux-3-image-klein-FP8, which carries the official diffusers folders
// byte-identical beside the FP8 / INT8 checkpoints. The GGUFs pair with the same repo for the Qwen3-VL
// text encoder and the FLUX.2 VAE.
@@ -289,22 +315,6 @@ export const IMAGE_CATALOG: CatalogGroup[] = [
gguf("unsloth/flux-3-image-klein-base-GGUF"),
],
},
- {
- // MiniMax-H3-Image: text to image and edits with 1-16 <Picture N> references. Size is the released bf16 total
- // (the backend quantizes at load); tiers are measured offloaded runs. CUDA / ROCm only.
- canonicalId: "MiniMaxAI/MiniMax-H3-Image",
- displayName: "MiniMax-H3-Image",
- description: "Text-to-image and image editing with up to 16 pictures",
- scope: "image",
- artifacts: [
- bf16Pipeline("MiniMaxAI/MiniMax-H3-Image", 126, {
- offloadFitTiers: [
- { gpuGb: 22, systemRamGb: 48, requiresDenseQuant: true },
- { gpuGb: 39, systemRamGb: 64, requiresDenseQuant: true },
- ],
- }),
- ],
- },
{
// MOSS-Image SFT: the bf16 row keeps the OpenMOSS id but fetches from unsloth/MOSS-Image-SFT-FP8 (upstream
// folders mirrored byte-identical). GGUFs are sd.cpp only, VAE and tokenizer from the same repo.
diff --git a/studio/frontend/tests/picker-auto-quant-text-encoder.test.ts b/studio/frontend/tests/picker-auto-quant-text-encoder.test.ts
index b828230cc0a..a35dde4965a 100644
--- a/studio/frontend/tests/picker-auto-quant-text-encoder.test.ts
+++ b/studio/frontend/tests/picker-auto-quant-text-encoder.test.ts
@@ -18,6 +18,7 @@ const KLEIN_BASE = "black-forest-labs/flux-3-image-klein-base";
const QWEN_21 = "unsloth/Qwen-Image-2.1";
const QWEN_21_TURBO = "Qwen/Qwen-Image-2.1-Turbo";
const KONTEXT = "unsloth/FLUX.1-Kontext-dev";
+const MINIMAX_H3_IMAGE = "MiniMaxAI/MiniMax-H3-Image";
const notDownloaded = () => false;
const GIB = 1024 ** 3;
@@ -48,6 +49,7 @@ test("each auto-quantised row states the hosted encoder file the backend loads",
[QWEN_21, 17_534_340_584, 9_349_769_248],
[QWEN_21_TURBO, 17_534_340_584, 9_349_769_248],
[KONTEXT, 9_524_648_584, 6_163_991_880],
+ [MINIMAX_H3_IMAGE, 66_726_510_529, 27_141_342_152],
] as const) {
const te = artifactForRepoId(id, IMAGE_CATALOG)?.artifact.autoQuantTextEncoderGb;
assert.ok(te, id);
diff --git a/studio/frontend/tests/recommended-catalog-seeds.test.ts b/studio/frontend/tests/recommended-catalog-seeds.test.ts
index 5e847025087..8dc1e6779f2 100644
--- a/studio/frontend/tests/recommended-catalog-seeds.test.ts
+++ b/studio/frontend/tests/recommended-catalog-seeds.test.ts
@@ -578,6 +578,7 @@ test("pinned families lead Recommended whatever the listing sort, lead rows firs
g.canonicalId.toLowerCase(),
);
assert.deepEqual(pinnedFamilies, [
+ "minimaxai/minimax-h3-image",
"black-forest-labs/flux-3-image-klein",
"unsloth/qwen-image-2.1",
"qwen/qwen-image-2.1-turbo",
@@ -585,7 +586,8 @@ test("pinned families lead Recommended whatever the listing sort, lead rows firs
const pinnedTogether = IMAGE_CATALOG.filter((g) => g.pinToTop && g.pinRowsTogether).map((g) =>
g.canonicalId.toLowerCase(),
);
- assert.deepEqual(pinnedTogether, ["black-forest-labs/flux-3-image-klein"]);
+ assert.deepEqual(pinnedTogether, ["minimaxai/minimax-h3-image", "black-forest-labs/flux-3-image-klein"]);
+ const h3i = "MiniMaxAI/MiniMax-H3-Image";
const flux3 = "black-forest-labs/flux-3-image-klein";
const flux3Gguf = "unsloth/flux-3-image-klein-GGUF";
const flux3Base = "black-forest-labs/flux-3-image-klein-base";
@@ -597,6 +599,7 @@ test("pinned families lead Recommended whatever the listing sort, lead rows firs
const qwen2512 = "unsloth/Qwen-Image-2512-GGUF";
const seeds: Row[] = [
{ id: zImageTurbo, isGguf: true },
+ { id: h3i },
{ id: flux3 },
{ id: flux3Gguf, isGguf: true },
{ id: flux3Base },
@@ -641,10 +644,12 @@ test("pinned families lead Recommended whatever the listing sort, lead rows firs
flux3Gguf,
turbo,
flux3,
+ h3i,
flux3Base,
]);
// Interleaved pins: every lead row ("Fast FP8") first, then every GGUF row.
assert.deepEqual(order(pinnedFamilies), [
+ h3i,
flux3,
qwen21,
turbo,
@@ -655,15 +660,14 @@ test("pinned families lead Recommended whatever the listing sort, lead rows firs
zImageTurbo,
flux3Base,
]);
- // A device the pipelines do not fit: the GGUF rows lead, still FLUX.3 first.
+ // A device the pipelines do not fit: the GGUF rows lead, still FLUX.3 first (MiniMax-H3-Image has no GGUF).
assert.deepEqual(
order(pinnedFamilies, (r) => Boolean(r.isGguf)),
[flux3Gguf, qwen21Gguf, turboGguf, qwen2512, zImageTurbo],
);
- // As shipped: "FLUX.3 Image Klein (Fast FP8)" on top with its GGUF right under it, then "Qwen-Image 2.1
- // (Fast FP8)" and "Qwen-Image 2.1 Turbo (Fast FP8)", then their GGUF rows. The unpinned Base keeps its
- // place among the other models.
+ // The unpinned Base keeps its place among the other models.
assert.deepEqual(order(pinnedFamilies, () => true, pinnedTogether), [
+ h3i,
flux3,
flux3Gguf,
qwen21,