# mlx-workspace

A workspace for running and benchmarking local LLMs on Apple Silicon with MLX, built to measure which local model can handle routine coding operations well enough to replace paid cloud requests in an internal coding tool.

mlx-workspace runs large language models locally on Apple Silicon using MLX and connects AI coding tools to that local server so their actual capability on real tasks can be measured, rather than assumed. Its purpose is answering one question for its parent project, nav-pilot: which local model is good enough to absorb routine coding operations that would otherwise consume a paid Copilot request.

A generated manifest lists the models currently shipped to nav-pilot users, each with its context size, sampling and prompt-cache tuning, memory requirements, minimum supported client version, and the specific task classes it has been verified to handle, with evidence linked to written decision records. Models that do not yet meet the bar are parked rather than shipped, and every task class not yet verified against a given model continues to run on the cloud model instead.

Website: https://github.com/navikt/mlx-workspace