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
Unsloth Studio: list a GGUF with no header metadata under On Device on the Images page (#11830)
Leo Borcherding4c45445042d8
diff --git a/studio/backend/hub/services/models/catalog_classification.py b/studio/backend/hub/services/models/catalog_classification.pyindex 5824e15e83d..a5d39af0c57 100644--- a/studio/backend/hub/services/models/catalog_classification.py+++ b/studio/backend/hub/services/models/catalog_classification.py@@ -164,16 +164,23 @@ def _unhydrated_gguf_task(name_hints: tuple[Optional[str], ...]) -> Optional[str"""if any(_is_h3_bundle_gguf_hint(hint) for hint in name_hints):return _VIDEO_GEN_TASK+ from core.inference.video_families import detect_video_family+# Leaves only: family detection matches a keyword in ANY path segment, so a chat GGUF under# .../FLUX.1-dev-GGUF/extra/ read as text-to-image.- return _name_hint_media_task(tuple(_hint_leaf(hint) for hint in name_hints if hint), None)+ leaves = tuple(_hint_leaf(hint) for hint in name_hints if hint)+ # H3 denoisers matched by prefix above; any other H3 name (qwen3vl_32b_minimax_h3-*) is the conditioner.+ if any(getattr(detect_video_family(leaf), "name", None) == "minimax-h3" for leaf in leaves):+ return None+ return _name_hint_media_task(leaves, None)def _arch_to_task(arch: Optional[str], name_hints: tuple[Optional[str], ...] = ()) -> Optional[str]:if any(_is_h3_bundle_gguf_hint(hint) for hint in name_hints):return _VIDEO_GEN_TASKif arch is None:- return None+ # Qwen-Image-2.1 GGUFs have kv_count 0, so the name is the only evidence, as for a cloud placeholder.+ return _unhydrated_gguf_task(name_hints)normalized = arch.lower()if normalized == "qwen3" and any(_QWEN3_ASR_HINT.search(str(hint)) for hint in name_hints if hint@@ -312,7 +319,7 @@ def _gguf_folder_task(complete = Falsecontinueif task is None:- # A truncated header gives no architecture, and _arch_to_task answers None.+ # No architecture and a name that says nothing: unclassified.complete = Falsecontinueif task in _LOADABLE_MEDIA_GGUF_TASKS:diff --git a/studio/backend/hub/tests/test_model_services.py b/studio/backend/hub/tests/test_model_services.pyindex 7b0ae731bb6..3531f35899b 100644--- a/studio/backend/hub/tests/test_model_services.py+++ b/studio/backend/hub/tests/test_model_services.py@@ -4070,6 +4070,32 @@ def test_qwen3_asr_gguf_name_hint_is_not_classified_as_chat(monkeypatch, tmp_patassert catalog_classification._gguf_path_task(chat) == "text-generation"+def test_a_gguf_with_no_architecture_is_classified_from_its_name(monkeypatch, tmp_path):+ """``unsloth/Qwen-Image-2.1-GGUF`` files have kv_count 0; a name that says nothing stays None."""+ from hub.services.models import catalog_classification++ image = tmp_path / "qwen-image-2.1-Q4_K_M.gguf"+ chat = tmp_path / "Some-Chat-7B-Q4_K_M.gguf"+ image.write_bytes(b"gguf")+ chat.write_bytes(b"gguf")+ monkeypatch.setattr(catalog_classification, "_gguf_architecture", lambda _path: None)+ monkeypatch.setattr(catalog_classification, "_gguf_family_buildable", lambda _hints: True)++ assert (+ catalog_classification._gguf_path_task(image, ("unsloth/Qwen-Image-2.1-GGUF",))+ == "text-to-image"+ )+ assert catalog_classification._gguf_path_task(chat) is None++ # H3's conditioner is kv_count 0 too; only the fl2va / ref2va denoisers are video.+ conditioner = tmp_path / "qwen3vl_32b_minimax_h3-Q4_K_M.gguf"+ denoiser = tmp_path / "minimax_h3_fl2va_pruned-Q4_K.gguf"+ conditioner.write_bytes(b"gguf")+ denoiser.write_bytes(b"gguf")+ assert catalog_classification._gguf_path_task(conditioner) is None+ assert catalog_classification._gguf_path_task(denoiser) == "text-to-video"++def test_local_inventory_filters_embedder_configured_by_snapshot_path(monkeypatch, tmp_path):from core.rag import config as rag_config