| Architecture | model_type |
Examples | Chat Format | Vision |
|---|---|---|---|---|
| Gemma 4 | gemma4, gemma4_unified |
gemma-4-e2b-it-4bit, gemma-4-e4b-it-8bit, gemma-4-26b-a4b-it-4bit, gemma-4-12b-unified |
Gemma turns | SigLIP (unified adds audio) |
| Gemma 3 | gemma3 |
gemma-3-12b-it-qat-4bit |
Gemma turns | -- |
| DiffusionGemma | diffusion_gemma |
diffusiongemma-26B-A4B-it-4bit |
Gemma turns (block diffusion) | -- |
| Qwen 2 / 3 / 3.5 / 3.6 / 3.8 | qwen2, qwen3, qwen3_moe, qwen3_5, qwen3_5_moe, qwen3_next |
Qwen3-4B, Qwen3.5-4B, Qwen3.6-27B, Qwen3.6-35B-A3B, Qwen3.8-27B (18.2 GB, draft head baked in, effort levels xhigh/medium/low) |
ChatML | Qwen3-VL |
| MiMo V2.6 Flash | mimo_v2, mimo_v2_flash |
Xiaomi MiMo-V2.6-Flash (309B-A15B): mlx-community/MiMo-V2.6-Flash-RL-mxfp4-q8 and the other mlx-lm-layout packs, or the MOPD release via tests/convert_mimo_v2.py (158 GB: the release's MXFP4 experts unchanged, 8-bit trunk) |
ChatML + XML tool calls, thinking default on | -- |
| Muse-Glimmer | muse_glimmer |
Meta's Muse-Glimmer-30B (4-bit / 8-bit, DFlash draft companion built in, up to 75 tok/s on M4 Max) | Harmony channels + ATEM tools | Muse ViT (images) |
| Ling 3.0 | bailing_hybrid |
inclusionAI Ling 3.0, e.g. rapid-mlx/Ling-3.0-tiny-MLX-4bit (4.2 GB, KDA + MLA hybrid MoE) |
GLM tags, thinking default on | -- |
| DeepSeek V4 Flash | deepseek_v4 |
DeepSeek-V4-Flash-0731 (284B-A13B, 1M ctx) — native MLX for safetensors builds, embedded ds4 for .gguf |
DSV4 + DSML tools | -- |
| DeepSeek V4.1 Flash | deepseek_v41 |
DeepSeek-V4.1-Flash: the MLX packs (pipenetwork's REAP50, Jundot's oQ) and the OpensourceWTF EXL3 streaming repack on mlx-stream (experts stream from SSD; a 128 GB Mac serves it). Engram tables stay on disk | V4.1 DSML tools, thinking default on, effort low/medium/high/max | -- |
| Inkling Small | inkling_mm_model |
Thinking Machines Inkling Small (276B-A12B MoE, 2-bit) | role-less channel messages | -- |
| Hunyuan 3 | hy_v3 |
Hy3-295B-Instruct (295B-A21B MoE, 2-bit) |
Hunyuan tags | -- |
| Laguna | laguna |
poolside Laguna S 2.1 / XS (117.6B-A8.5B MoE coder, nvfp4) | GLM tags, pre-opened think | -- |
| Kolibri-1 | kolibri1 |
here-be-dragons-ai/Kolibri-1-MLX-3bit (78B-A3.5B MoE, 33 GiB) |
ChatML, reasoning_effort |
-- |
| Nemotron-H | nemotron_h |
Nemotron-3-Nano-4B | ChatML | -- |
| LFM2 / LFM2.5 | lfm2, lfm2_vl |
LFM2.5-2.6B (8-bit, bf16, nvfp4, mxfp4), LFM2.5-VL 3B / 1.6B | ChatML, Pythonic tool calls | SigLIP2, big images tiled |
| Llama | llama |
Llama 3, Llama 3.1, Llama 3.2 | Llama-3 | -- |
| Mistral | mistral |
Mistral 7B Instruct v0.3 | Mistral turns | -- |
| Laya (typed decisions) | laya |
aac6fef/laya-multilingual-mlx (mmBERT-base encoder + decision head; POST /v1/decisions) |
n/a | -- |
| Kev (typed decisions) | kev |
aselea/Kev-4B-MLX-Serve-8bit (Qwen3.5-4B + pointer head; POST /v1/decisions; packs from tests/convert_kev_weights.py) |
n/a | -- |
| Embeddings | bert, gemma3_text, embedding_gemma2, qwen3 |
bge, mxbai, EmbeddingGemma 1 and 2 (2: 768 dims, dimensions down to 128, 8192 tokens, text, image and video), Qwen3-Embedding (pooling read from the checkpoint) |
n/a | EmbeddingGemma 2: SigLIP (images, video) |
| Anything else as GGUF | via embedded llama.cpp | any .gguf on HuggingFace |
per-template | -- |
Media models live in the same registry and are classified the same way: FLUX.2, Krea-2, Mage-Flow and Qwen-Image-2.1 (image), Qwen3-TTS, Kokoro, ACE-Step and MiniMax Music 3 (speech + music), LTX-Video 2.3 / 2.5 and MiniMax-H3 (video), Hunyuan3D-2.1 (3D). A chat request naming one of them gets a 400 that names the endpoint to use instead.
Any quantized MLX model using one of the above architectures works natively. Anything else can be served as GGUF through the embedded llama.cpp engine — just pick the .gguf file in the Model Browser and the server auto-routes by format. Models with unsupported architectures are flagged in the Model Browser but can still be downloaded.