Models
The model classes the engine handles and the families it was checked with.
Any architecture mlx-lm loads runs as text; the table lists what the engine adds per family, with one model standing for each. A mark is built, a ring is planned, a dash means the family has no such thing. Hover a mark for what it stands for. Every class has a tiny model in the test suite; measured numbers are published as they are measured, with machine, model and date.
| Family | Architecture | Thinking | Tool calls | Vision | Draft head |
|---|---|---|---|---|---|
| Gemma 4google/gemma-4-12B-it, E-series | |||||
| GLM-4.7zai-org/GLM-4.7-Flash | — | ||||
| gpt-ossopenai/gpt-oss-20b | — | — | |||
| Granite 4ibm-granite/granite-4.0-h-small | — | — | |||
| Granite Vision 4.1ibm-granite/granite-vision-4.1-4b | — | — | |||
| Mistral Small 3.xmistralai/Mistral-Small-3.2-24B-Instruct-2506 | — | — | |||
| Muse Glimmermeta-models/Muse-Glimmer-30B | |||||
| Qwen3.5 / 3.6 / 3.8Qwen/Qwen3.8-27B | |||||
| Qwen3.8 Flash-NextQwen/Qwen3.8-Flash-Next |
Tool parsers vendored from mlx-lm cover more families than the table (Kimi, MiniMax, LongCat, pythonic and JSON calls), and think markers are read from the template of any model.