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.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Studio: pin FLUX.3 Image Klein (Fast FP8) to the top of the Images picker (#14184)
Daniel Hanb2276549274c
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.tsindex 313dc024289..4821ce7c0c7 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@@ -158,6 +158,35 @@ assert.equal(groupForRepoId("unsloth/Qwen-Image-2.1-GGUF", IMAGE_CATALOG)?.canon"Qwen-Image-2.1-Turbo-GGUF (Slow)",);}+// FLUX.3 Klein: a dense row on the gated BFL id (fetched from the FP8 repo), so it reads "Fast FP8" like 2.1.+for (const [base, fp8Repo, name] of [+ ["black-forest-labs/flux-3-image-klein", "unsloth/flux-3-image-klein-FP8", "FLUX.3 Image Klein"],+ ["black-forest-labs/flux-3-image-klein-base", "unsloth/flux-3-image-klein-base-FP8", "FLUX.3 Image Klein Base"],+] as const) {+ const group = groupForRepoId(base, IMAGE_CATALOG);+ assert.ok(group, base);+ assert.equal(groupForRepoId(fp8Repo, IMAGE_CATALOG), group, fp8Repo);+ assert.equal(groupForRepoId(`unsloth/${base.split("/")[1]}-GGUF`, IMAGE_CATALOG), group, base);+ assert.equal(curatedRowLabelFor(base, IMAGE_CATALOG, "dense-quant", ["int8", "fp8"])?.name, `${name} (Fast FP8)`);+ assert.equal(+ curatedRowLabelFor(`unsloth/${base.split("/")[1]}-GGUF`, IMAGE_CATALOG, "dense-quant", ["fp8"])?.name,+ `${base.split("/")[1]}-GGUF (Slow)`,+ );+}+assert.equal(+ IMAGE_CATALOG.find((g) => g.pinToTop)?.canonicalId,+ "black-forest-labs/flux-3-image-klein",+);+// The pinned Fast FP8 row must survive "Fit on device" on a 24 GB card (6.0 GiB DiT + 8.6 GiB companions, under+// the 70% rule), else the Slow GGUF row takes the top. A host without the dense quant sizes it in bf16 and drops it.+for (const id of ["black-forest-labs/flux-3-image-klein", "black-forest-labs/flux-3-image-klein-base"]) {+ assert.equal(+ curatedArtifactFitsDevice(id, IMAGE_CATALOG, { gpuGb: 24, systemRamGb: 64, denseQuantSchemes: ["int8", "fp8"] }),+ true,+ id,+ );+ assert.equal(curatedArtifactFitsDevice(id, IMAGE_CATALOG, { gpuGb: 24, systemRamGb: 64 }), false, id);+}assert.notEqual(groupForRepoId("Qwen/Qwen-Image", IMAGE_CATALOG), qwen21);assert.equal(groupForRepoId("Qwen/Qwen-Image", IMAGE_CATALOG)?.canonicalId,@@ -1259,6 +1288,8 @@ const PREQUANT_ROWS = [["krea/Krea-2-Turbo", "unsloth/Krea-2-Turbo-FP8"],["Qwen/Qwen-Image-2.1", "unsloth/Qwen-Image-2.1-FP8"],["Qwen/Qwen-Image-2.1-Turbo", "unsloth/Qwen-Image-2.1-Turbo-FP8"],+ ["black-forest-labs/flux-3-image-klein", "unsloth/flux-3-image-klein-FP8"],+ ["black-forest-labs/flux-3-image-klein-base", "unsloth/flux-3-image-klein-base-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.tsindex b2595dc4064..f02135b5607 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@@ -236,6 +236,47 @@ export const IMAGE_CATALOG: CatalogGroup[] = [scope: "image",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" row leads the list above Qwen-Image 2.1's. 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.+ canonicalId: "black-forest-labs/flux-3-image-klein",+ displayName: "FLUX.3 Image Klein",+ description: "Text-to-image and image editing",+ scope: "image",+ // Same reason as Qwen-Image 2.1's alias: the int8 half of the prequant repo has no artifact row.+ aliases: ["unsloth/flux-3-image-klein-FP8"],+ pinToTop: true,+ artifacts: [+ // Resident bf16 GiB, as the fit rule reads it: 10.34 DiT + 8.27 Qwen3-VL-4B + 0.31 VAE.+ bf16Pipeline("black-forest-labs/flux-3-image-klein", 18.92, {+ totalParams: 5548681984,+ prequantRepo: "unsloth/flux-3-image-klein-FP8",+ prequantSizeGb: { fp8: 6.0, int8: 6.0 },+ }),+ gguf("unsloth/flux-3-image-klein-GGUF"),+ ],+ },+ {+ // The undistilled checkpoint (CFG, more steps), laid out the same way from its own FP8 repo. Not pinned:+ // it sorts with the other models.+ canonicalId: "black-forest-labs/flux-3-image-klein-base",+ displayName: "FLUX.3 Image Klein Base",+ description: "Text-to-image and image editing",+ scope: "image",+ aliases: ["unsloth/flux-3-image-klein-base-FP8"],+ artifacts: [+ bf16Pipeline("black-forest-labs/flux-3-image-klein-base", 18.9, {+ totalParams: 5538458368,+ prequantRepo: "unsloth/flux-3-image-klein-base-FP8",+ // The INT8 file is BFL's own int8_convrot denoiser, smaller than the FP8.+ prequantSizeGb: { fp8: 5.98, int8: 5.41 },+ }),+ gguf("unsloth/flux-3-image-klein-base-GGUF"),+ ],+ },{canonicalId: "unsloth/Qwen-Image-2.1",displayName: "Qwen-Image 2.1",@@ -359,23 +400,6 @@ export const IMAGE_CATALOG: CatalogGroup[] = [scope: "image",artifacts: [gguf("unsloth/FLUX.2-klein-9B-GGUF")],},- {- // FLUX.3 Image Klein (distilled): text to image and up to 10 reference images. The GGUFs pair with the- // official repo for the Qwen3-VL text encoder and the FLUX.2 VAE.- canonicalId: "black-forest-labs/flux-3-image-klein",- displayName: "FLUX.3 Image Klein",- description: "Text-to-image and image editing",- scope: "image",- artifacts: [gguf("unsloth/flux-3-image-klein-GGUF")],- },- {- // The undistilled checkpoint (CFG, more steps); same companions from its own official repo.- canonicalId: "black-forest-labs/flux-3-image-klein-base",- displayName: "FLUX.3 Image Klein Base",- description: "Text-to-image and image editing",- scope: "image",- artifacts: [gguf("unsloth/flux-3-image-klein-base-GGUF")],- },{canonicalId: "unsloth/Qwen-Image-Edit-2511",displayName: "Qwen-Image-Edit 2511",diff --git a/studio/frontend/tests/recommended-catalog-seeds.test.ts b/studio/frontend/tests/recommended-catalog-seeds.test.tsindex 93a4ee46683..4af52dbd584 100644--- a/studio/frontend/tests/recommended-catalog-seeds.test.ts+++ b/studio/frontend/tests/recommended-catalog-seeds.test.ts@@ -624,7 +624,14 @@ test("pinned families lead Recommended whatever the listing sort, lead rows firsconst pinnedFamilies = IMAGE_CATALOG.filter((g) => g.pinToTop).map((g) =>g.canonicalId.toLowerCase(),);- assert.deepEqual(pinnedFamilies, ["unsloth/qwen-image-2.1", "qwen/qwen-image-2.1-turbo"]);+ assert.deepEqual(pinnedFamilies, [+ "black-forest-labs/flux-3-image-klein",+ "unsloth/qwen-image-2.1",+ "qwen/qwen-image-2.1-turbo",+ ]);+ 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";const qwen21 = "unsloth/Qwen-Image-2.1";const qwen21Gguf = "unsloth/Qwen-Image-2.1-GGUF";const turbo = "Qwen/Qwen-Image-2.1-Turbo";@@ -633,18 +640,22 @@ test("pinned families lead Recommended whatever the listing sort, lead rows firsconst qwen2512 = "unsloth/Qwen-Image-2512-GGUF";const seeds: Row[] = [{ id: zImageTurbo, isGguf: true },+ { id: flux3 },+ { id: flux3Gguf, isGguf: true },+ { id: flux3Base },{ id: qwen21 },{ id: qwen21Gguf, isGguf: true },{ id: turbo },{ id: turboGguf, isGguf: true },{ id: qwen2512, isGguf: true },];- // Qwen-Image 2.1 and its Turbo trend last, so only the pin lifts them.+ // FLUX.3 Klein, Qwen-Image 2.1 and its Turbo trend last, so only the pin lifts them.const results: Row[] = [{ id: qwen2512, isGguf: true },{ id: zImageTurbo, isGguf: true },{ id: turboGguf, isGguf: true },{ id: qwen21Gguf, isGguf: true },+ { id: flux3Gguf, isGguf: true },];const order = (pinned?: readonly string[], fits: (r: Row) => boolean = () => true) =>ids(@@ -658,14 +669,35 @@ test("pinned families lead Recommended whatever the listing sort, lead rows firspinnedFamilies: pinned,}),);- // Unpinned, the listing sorts them and the vendor-hosted Turbo pipeline trails the first-party rows.- assert.deepEqual(order(), [qwen2512, zImageTurbo, turboGguf, qwen21, qwen21Gguf, turbo]);- // "Qwen-Image 2.1 Turbo (Fast FP8)" directly under "Qwen-Image 2.1 (Fast FP8)", then their GGUF rows.- assert.deepEqual(order(pinnedFamilies), [qwen21, turbo, qwen21Gguf, turboGguf, qwen2512, zImageTurbo]);- // A device the pipelines do not fit: the GGUF rows lead, still 2.1 first.+ // Unpinned, the listing sorts them and the vendor-hosted pipelines trail the first-party rows.+ assert.deepEqual(order(), [+ qwen2512,+ zImageTurbo,+ turboGguf,+ qwen21,+ qwen21Gguf,+ flux3Gguf,+ turbo,+ flux3,+ flux3Base,+ ]);+ // "FLUX.3 Image Klein (Fast FP8)" on top, 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.+ assert.deepEqual(order(pinnedFamilies), [+ flux3,+ qwen21,+ turbo,+ flux3Gguf,+ qwen21Gguf,+ turboGguf,+ qwen2512,+ zImageTurbo,+ flux3Base,+ ]);+ // A device the pipelines do not fit: the GGUF rows lead, still FLUX.3 first.assert.deepEqual(order(pinnedFamilies, (r) => Boolean(r.isGguf)),- [qwen21Gguf, turboGguf, qwen2512, zImageTurbo],+ [flux3Gguf, qwen21Gguf, turboGguf, qwen2512, zImageTurbo],);});