Machine learning built for Apple silicon
MLX is an open source framework from Apple’s machine learning research team for running machine learning efficiently on Apple silicon. It takes advantage of unified memory, where CPU and GPU share the same pool, and supports training, inference, and fine-tuning.
It can be used from Python, Swift, C, and C++, with related tools such as MLX-LM for language models and MLX Whisper for speech recognition. It is a common choice for trying local LLMs on a Mac.
When it appeared, and what changed
MLX is not new in WWDC26 — Apple’s machine learning research team released it in December 2023, and it has grown since into a base for local AI development on Apple silicon.
WWDC26 covered running models under MLX and reaching them from agents and Xcode through an OpenAI-compatible local server, as well as distributing a large model across several connected Macs.
What you can do
- Run LLMs and image models on a Mac
- Quantize a model to reduce the memory it needs
- Fine-tune a model on your own data
- Stand up a local AI server for apps and agents to call
- Distribute processing across several Apple silicon machines
How it differs from Core ML and the Foundation Models framework
MLX is a machine learning framework for research, experimentation, and model adjustment. The Foundation Models framework is the API for using the models built into the OS from a Swift app.
MLX lets you run a wide range of open models on a Mac, but the source, license, memory requirements, and safety of each model are yours to verify. Larger models need substantial unified memory and storage.