Meta released Muse Spark Open, a mixture-of-experts model with 816 billion total parameters and 42 billion active, under the Muse Model License. The weights are public on Hugging Face as Muse-Spark-Open-816B-A42B, with a 1 million-token context window. Meta is calling it "the most capable US open-weight model to date."
The interesting detail is not the parameter count but where Meta expects it to run. The release is optimised for two to four paired M5 Ultra Mac Studios via MLX — a deployment target that puts a frontier-class open model inside a desk-sized box rather than a rack. At 42B active parameters the compute cost per token is far below what 816B suggests, which is what makes that target plausible at all.
The licence is the thing to read before planning around it. Muse Model License is not Apache 2.0 or MIT; like previous Meta releases it carries its own terms, and anyone building commercially should check them against their use case rather than assuming open weights means unrestricted. Independent benchmark results were not available at publication.
Sources:
Meta Research ·
Hugging Face ·
Unite.AI