MLX LM

MLX LM is Apple’s own Python package for running, fine-tuning and deploying large language models on Apple Silicon with the MLX framework. It is the easiest official way to put models like Llama, Qwen, Mistral and Gemma on a Mac: install one package, name a model, and generate — using the unified memory of M-series chips efficiently. The project has earned about 7,400 stars on GitHub and is updated very actively in 2026.

Overview

The package covers the whole local-LLM workflow. Generation works from the command line or a Python API, with chat templating, streaming and an OpenAI-style server mode for wiring local models into other tools. For adaptation, it includes LoRA and full fine-tuning with straightforward configuration, so you can specialize a model on your own data on a single Mac. Models from Hugging Face can be converted to MLX format and quantized — typically to 4 bits — shrinking them dramatically while keeping quality high, and a large catalog of pre-converted models is ready to download. Because MLX is built by Apple’s machine learning research team specifically for Apple Silicon, performance per watt is excellent and memory is shared between CPU and GPU instead of copied. For Mac-based AI development, MLX LM is the reference implementation most other tools build on.

Key Features

• Run LLMs like Llama, Qwen, Mistral and Gemma natively on Apple Silicon
• Command-line and Python APIs, with streaming chat and a server mode
• LoRA and full fine-tuning on a single Mac
• Convert Hugging Face models to MLX and quantize them (e.g. 4-bit)
• Large catalog of pre-converted models ready to download
• Built by Apple’s ML research team on the MLX framework

💳 Pricing

MLX LM is free and open source under the MIT license. The models it runs are open weights.

Who Is It For

MLX LM is for Mac developers and researchers who want to run, serve and fine-tune open LLMs locally on Apple Silicon with first-party tooling.