Meta releases Muse Spark Open, an 816B open-weight model that runs on Mac Studios

Meta Superintelligence Labs released Muse Spark Open, the open-weight version of its Muse Spark model, and Meta calls it the most capable US open-weight model to date. It is an 816 billion parameter mixture-of-experts model with about 42 billion parameters active per token. It takes text, images, audio and video as input, returns text, and holds a 1 million token context.

It was trained for agentic coding: multi-file refactors, long debugging sessions and tool use. The surprise is where it runs. Meta worked with Apple so that a compressed NVFP4 build runs locally on two M5 Ultra Mac Studios with 512GB of memory each, using Apple's updated MLX toolkit. Meta says that is a first for a model this size, and it showed the model fixing an iOS app and editing video without the cloud.

On the Artificial Analysis Intelligence Index it scores 39, ahead of other Western open-weight models but behind China's GLM 5.3 at 45. It is already available on major cloud platforms and in common open-source tooling. The license asks users to credit the model as Built with Muse Spark Open.

Our take

The headline is the deployment, not the benchmark. Frontier-size open weights running on a desk changes who can afford private, local AI for serious coding work.