Catalog

Supported models

Ungated, single-file GGUFs verified against the Hugging Face API and run through llama.cpp; vision models ship a projector file alongside; Whisper models are whisper.cpp ggml files. Tools means llama.cpp has a native tool-call parser for the family; Vision means the model takes image input. Every entry is one omg model pull away.

V0.4.0 · GGUF + GGML · LLAMA.CPP + WHISPER.CPP

Family

IDFamilySource repoQuantSizeToolsVisionPull

smollm2-135m-instructSmolLM2bartowski/SmolLM2-135M-Instruct-GGUFQ8_00.14 GB——

qwen2.5-0.5b-instructQwen2.5Qwen/Qwen2.5-0.5B-Instruct-GGUFQ4_K_M0.49 GB●—

qwen2.5-1.5b-instructQwen2.5Qwen/Qwen2.5-1.5B-Instruct-GGUFQ4_K_M1.1 GB●—

qwen2.5-3b-instructQwen2.5Qwen/Qwen2.5-3B-Instruct-GGUFQ4_K_M2.1 GB●—

qwen2.5-7b-instructQwen2.5bartowski/Qwen2.5-7B-Instruct-GGUFQ4_K_M4.7 GB●—

qwen3-4b-instructQwen3unsloth/Qwen3-4B-Instruct-2507-GGUFQ4_K_M2.5 GB●—

llama-3.2-1b-instructLlama 3bartowski/Llama-3.2-1B-Instruct-GGUFQ4_K_M0.81 GB●—

llama-3.2-3b-instructLlama 3bartowski/Llama-3.2-3B-Instruct-GGUFQ4_K_M2.0 GB●—

llama-3.1-8b-instructLlama 3bartowski/Meta-Llama-3.1-8B-Instruct-GGUFQ4_K_M4.9 GB●—

phi-4-mini-instructPhi-4bartowski/microsoft_Phi-4-mini-instruct-GGUFQ4_K_M2.5 GB——

gemma-3-1b-itGemma 3ggml-org/gemma-3-1b-it-GGUFQ4_K_M0.81 GB——

gemma-3-4b-itGemma 3ggml-org/gemma-3-4b-it-GGUFQ4_K_M2.5 GB—●

gemma-3-12b-itGemma 3ggml-org/gemma-3-12b-it-GGUFQ4_K_M7.3 GB——

qwen2.5-vl-3b-instructQwen2.5-VLggml-org/Qwen2.5-VL-3B-Instruct-GGUFQ4_K_M1.9 GB—●

qwen2.5-vl-7b-instructQwen2.5-VLggml-org/Qwen2.5-VL-7B-Instruct-GGUFQ4_K_M4.7 GB—●

smolvlm-256m-instructSmolVLMggml-org/SmolVLM-256M-Instruct-GGUFQ8_00.18 GB—●

whisper-tinySPEECHWhisperggerganov/whisper.cppf160.08 GB——

whisper-baseSPEECHWhisperggerganov/whisper.cppf160.15 GB——

whisper-smallSPEECHWhisperggerganov/whisper.cppf160.49 GB——

whisper-large-v3-turboSPEECHWhisperggerganov/whisper.cppq8_00.87 GB——

20 of 20 models · verified 2026-08-15source: crates/core/src/catalog.rs → recipes/*.yaml

Beyond the catalog

Any GGUF, by reference

Anything the catalog doesn't carry can still be pulled — unsupported, but usually fine. Gated repos need a token via HF_TOKEN; add --mmproj for a vision model, or pull a whisper.cpp ggml-*.bin the same way.

$ omg model pull hf:bartowski/SmolLM2-360M-Instruct-GGUF/SmolLM2-360M-Instruct-Q4_K_M.gguf
$ omg model pull https://models.example.com/my-model.gguf --id my-model
$ omg model pull hf:owner/repo/ggml-medium.bin # a whisper model — kind inferred from the name

Coming

The catalog becomes recipes

The compiled-in catalog is migrating to recipes/*.yaml — one data file per model with quant variants, VRAM requirements, engine args and a smoke test. Adapting a new model becomes a pull request, not a code change.

Recipe spec →·Schema v1 announcement →