llm-mlx

Support for MLX models in LLM.

Read my blog for background on this project.

Installation

Install this plugin in the same environment as LLM. This plugin likely only works on macOS.

llm install llm-mlx

This plugin depends on sentencepiece which does not yet publish a binary wheel for Python 3.13. You will find this plugin easier to run on Python 3.12 or lower. One way to install a version of LLM that uses Python 3.12 is like this, using uv:

uv tool install llm --python 3.12

See issue #7 for more on this.

Usage

To install an MLX model from Hugging Face, use the llm mlx download-model command. This example downloads 1.8GB of model weights from mlx-community/Llama-3.2-3B-Instruct-4bit:

llm mlx download-model mlx-community/Llama-3.2-3B-Instruct-4bit

Then run prompts like this:

llm -m mlx-community/Llama-3.2-3B-Instruct-4bit 'Capital of France?' -s 'you are a pelican'

The mlx-community organization is a useful source for compatible models.

Models to try

The following models all work well with this plugin:

• mlx-community/Qwen2.5-0.5B-Instruct-4bit - 278MB
• mlx-community/Mistral-7B-Instruct-v0.3-4bit - 4.08GB
• mlx-community/Mistral-Small-24B-Instruct-2501-4bit — 13.26 GB
• mlx-community/DeepSeek-R1-Distill-Qwen-32B-4bit - 18.5GB
• mlx-community/Llama-3.3-70B-Instruct-4bit - 40GB

Model options

MLX models can use the following model options:

• -o max_tokens INTEGER: Maximum number of tokens to generate in the completion (defaults to 1024)
• -o unlimited 1: Generate an unlimited number of tokens in the completion
• -o temperature FLOAT: Sampling temperature (defaults to 0.8)
• -o top_p FLOAT: Sampling top-p (defaults to 0.9)
• -o min_p FLOAT: Sampling min-p (defaults to 0.1)
• -o min_tokens_to_keep INT: Minimum tokens to keep for min-p sampling (defaults to 1)
• -o seed INT: Random number seed to use

For example:

llm -m mlx-community/Llama-3.2-3B-Instruct-4bit 'Joke about pelicans' -o max_tokens 60 -o temperature 1.0

Importing existing models

If you have used MLX models in the past you may already have some installed in your ~/.cache/huggingface/hub directory.

The llm mlx import-models command can detect these and provide you with the option to add them to the list of models registered with LLM.

llm mlx import-models

This will open an interface like this one:

Available models (↑/↓ to navigate, SPACE to select, ENTER to confirm, Ctrl+C to quit):
> ○ (llama) mlx-community/DeepSeek-R1-Distill-Llama-8B (already imported)
  ○ (llama) mlx-community/Llama-3.2-3B-Instruct-4bit (already imported)
  ○ (llama) mlx-community/Llama-3.3-70B-Instruct-4bit
  ○ (mistral) mlx-community/Mistral-7B-Instruct-v0.3-4bit (already imported)
  ○ (mistral) mlx-community/Mistral-Small-24B-Instruct-2501-4bit

Navigate and , hit <space> to select models to import and then hit <enter> to confirm.

Using models from Python

If you have registered models with the llm download-model command you can use in Python like this:

import llm
model = llm.get_model("mlx-community/Llama-3.2-3B-Instruct-4bit")
print(model.prompt("hi").text())

You can avoid that registration step entirely by accessing the models like this instead:

from llm_mlx import MlxModel
model = MlxModel("mlx-community/Llama-3.2-3B-Instruct-4bit")
print(model.prompt("hi").text())
# Outputs: How can I assist you today?

The LLM Python API documentation has more details on how to use LLM models.

Development

To set up this plugin locally, first checkout the code. Then create a new virtual environment:

cd llm-mlx
python -m venv venv
source venv/bin/activate

Now install the dependencies and test dependencies:

llm install -e '.[test]'

To run the tests:

python -m pytest