Meta releases open weights for Muse Spark

Meta published open weights for Muse Spark Open, a mixture-of-experts model with 816 billion total parameters and 42 billion active per token, spread across 256 experts with eight routed for any given token. It takes text, images, audio and video as input and produces text, with a context window of one million tokens, and Meta built it for coding and long-horizon agentic work. Hugging Face, which hosts the weights, called it “the most capable US open-weight model to date.”

The release is the open counterpart to a closed line. Meta went closed with Muse Spark in April, reversing the open-weight approach it had used for Llama, then released a small open model, Muse Glimmer, in August, and a closed, agent-focused pair, Muse Wonder, three days before this release. Muse Spark Open gives the Muse family an open-weight frontier-scale entry alongside those closed releases rather than replacing them.

Meta said the model runs on two Apple Mac Studios fitted with the new M5 Ultra chip, using Apple’s MLX framework — the unified memory across the two machines, Meta said, lets them behave “like one supercomputer with a terabyte of unified memory”, putting a frontier-scale model within reach of a well-equipped desktop rather than only a data centre. The weights are also available through API platforms including Baseten, Fireworks, Groq and Together, and for download via Amazon SageMaker, Google Cloud Model Garden and Microsoft Foundry Models as well as Hugging Face. They are released under a custom “Muse Model License” rather than a standard open-source licence such as Apache 2.0 or MIT; Meta describes it as carrying “limited commercial terms” that let most developers use the model free of charge.

Meta’s own evaluation report compares the model against other large open-weight systems, including Reflection AI’s Beam and Mistral Large 4, as well as several Chinese releases, but presents the actual scores as a graphic rather than in the accompanying text, and the report notes that rivals’ numbers were often self-reported rather than independently verified because their weights were not yet broadly available for Meta to test directly. The release adds a second large Western open-weight model to the market within ten days, following Mistral’s and Reflection AI’s announcements of similarly sized open models earlier in October — a brief cluster of large open releases from labs backed by American and European capital, an area recent frontier-scale activity had mostly ceded to Chinese developers.