MLX-LM
Apple's MLX framework applied to language models: a toolkit for running, serving and fine-tuning LLMs natively on Apple Silicon, using the unified-memory architecture rather than a CUDA path ported across. It includes quantisation and LoRA fine-tuning, and underpins the MLX model builds that apps like LM Studio and Ollama can now use on a Mac.
Strengths
- Runs thousands of Hugging Face models on Apple Silicon with a single command
- LoRA, QLoRA, and full fine-tuning, including on quantised models
- Built-in quantisation and model uploading
Weaknesses
- Apple Silicon only: no cross-platform support
- Models large relative to the machine memory run slowly
- Some tokenizers need trust_remote_code, adding friction
At a glance
- Licence
- MIT
- Pricing
- Free
- Platforms
- macOS
- GPU required
- No
Links
Works with
Glossary
Entry last verified 18 August 2026.