Updated October 2026

Run a local AI agent on Mac with MLX

MLX is Apple's machine-learning framework for Apple Silicon, and MLX-LM runs language models on it efficiently. CoWork OS can start an MLX-LM server for you and use it as a fully local model for agent tasks: files, tools, schedules and chat-app replies, with nothing sent to a cloud model.

Requirements

A Mac with Apple Silicon running macOS 13 or later, with the app running natively (not under Rosetta).
python3 available to the app.
The mlx-lm Python package.
Enough memory for your model. There is no fixed minimum; start with a smaller 4-bit model and move up.

1. Install MLX-LM

python3 -m pip install --upgrade mlx-lm
python3 -c "import mlx_lm; import mlx.core; print('MLX-LM ready')"

2. Start it from CoWork OS

Install CoWork OS and open Settings → AI & Models → Model Access.
Choose MLX (Apple Silicon) and pick a model. Leave the API key empty.
Click Start MLX Server and wait for Server running. The first run downloads the model from Hugging Face.
Click Test Connection.

Behind the scenes CoWork OS runs python3 -m mlx_lm.server --model <model> --port 8080 and connects to http://localhost:8080/v1.

Recommended models

Model	Notes
`mlx-community/Qwen3-8B-4bit`	A good starting point on most Apple Silicon Macs
`mlx-community/Qwen3-14B-4bit`	Better quality if you have more memory
`mlx-community/Qwen3-30B-A3B-4bit`	Mixture-of-experts model; strong for its speed
`mlx-community/Qwen3.6-35B-A3B-4bit-DWQ`	For Macs with plenty of unified memory

MLX or Ollama?

Both run models locally. MLX is built specifically for Apple Silicon; Ollama also runs on Windows and Linux and has a large model library. CoWork OS supports both, plus oMLX if you already run an oMLX server.

The full setup reference, including troubleshooting, is in the MLX-LM documentation.

Frequently asked questions

Can I run an AI agent locally on a Mac?

Yes. On an Apple Silicon Mac with macOS 13 or later, CoWork OS can start an MLX-LM server and use it as a local model for agent tasks.

How much RAM do I need for MLX?

It depends on the model. Start with a smaller 4-bit model such as Qwen3-8B-4bit and move to larger models if your Mac has more unified memory.

Does MLX work on Intel Macs?

No. MLX requires Apple Silicon, and the app must run natively rather than under Rosetta.

Is MLX faster than Ollama on a Mac?

MLX is designed for Apple Silicon and often performs well there, but speed depends on the model and quantization. CoWork OS supports both, so you can try each with the same task.

Install CoWork OS free Run your first task

CoWork OS itself is free and open source. Model providers, gateways, connected services, and account plans may charge separately and apply their own eligibility and usage rules.

Last reviewed 2026-10-20 against official documentation. Features, plans and prices change; check the linked sources before deciding.