What is laguna-mlx?

This repository adapts a large AI model called Laguna, made by a company called Poolside, to run on Apple Silicon Mac computers. The original Laguna model is designed for agentic coding tasks and can work with very long passages of text, up to one million words or tokens at once. The project provides the necessary code to run this model using Apple's MLX framework, which is built for Apple chips. A single Python file contains the model definition, and you register it once before running prebuilt versions of the model that are hosted online. The model uses a design called Mixture of Experts, where it has 256 separate specialist components but only activates about 10 of them for any given piece of text. This keeps the active computing load small, around 8 billion parameters, even though the total model is much larger at 118 billion. The architecture mixes two types of attention across its 48 layers, with some layers looking at the full text and others using a sliding window that focuses on nearby words. It also includes a shared expert that processes every token, plus specific mathematical techniques for handling position and gating. To use it, you install the MLX library and register the model file. The project offers several sizes, from a 2-bit compressed version at 35 gigabytes up to a full precision version at 219 gigabytes. The 4-bit version at 62 gigabytes is listed as the balanced default, and you need a Mac with at least 96 gigabytes of unified memory to run it comfortably. You can generate text from the command line, use a Python script, or run a local server that accepts requests compatible with OpenAI's format. The repository includes scripts to download the original weights, convert them into the different compressed sizes, and upload them. The 2-bit version uses a mixed precision approach because pure 2-bit compression produces incoherent output. By keeping key parts at 4-bit and only compressing the specialist components to 2-bit, the result stays coherent while staying compact. Tests verify the model structure against the original without requiring the full weight download. The code is under an Apache 2.0 license, while the model weights follow a separate license inherited from the original model.