Models¶
One page per model family: how it was converted, how to run it, and what was measured. To pick a download by how much memory your Mac has, start with Choose a model.
| Family | Type | Pages cover |
|---|---|---|
| GPT-OSS | MoE, 20B and 120B | 2-, 3- and 4-bit builds; the 48 GB hybrid; 131K context on 64 GB |
| Qwen3.5-122B-A10B | MoE, 256 experts | 3-bit and ternary; streaming on a 16 GB mini |
| Qwen3.6-35B-A3B | Hybrid MoE | Resident and streamed; the 16 GB ternary and agentic down4 builds |
| Qwen3.6-27B | Dense hybrid coder | A local coding model for a 48 GB Mac |
| Qwen3-235B-A22B | MoE, 128 experts | Converting on a 16 GB mini; hybrid, 3-bit and ternary builds |
| Nemotron-3 | Mamba/attention hybrid | Nano-4B and Super-120B |
| Nemotron 3.5 Super VL | Hybrid MoE vision model | 120B resident on a 64 GB Mac |
| Poolside Laguna | MoE coding agents | XS.2 33B and S-2.1 118B |
| Kimi K3 | 2.8T MoE | Streaming from disk on a 512 GB Mac Studio |
| Sarvam MoE | MoE | Port only; no build published |
| Muse Glimmer | Dense vision model, 30B | 4-bit and 3-bit builds that beat affine 4-bit |
| Qwen3.8-27B | Dense hybrid vision model | A 16 GB build and a 4-bit build |
| Qwen3.8-Flash-Next | MoE, 180B, 512 experts | Resident on a 64 GB Mac; n-gram offload |
| DiffusionGemma | Block-diffusion MoE | Experimental; a 16 GB build |
Every published build is in the TurboQuant-MLX collection on Hugging Face, and each card has the full measurements.