Meta said on October 15 that it has published the weights of Muse Spark Open, a model for coding and agent work. Weights are the trained numbers that let others run a model themselves. Meta calls it the strongest open-weight model made in the US so far. It has 816 billion parameters, with 42 billion active, and a 1 million-token context window. Downloads are on Hugging Face and three big cloud platforms.
Meta's account of the hardware is simple. Each Apple M5 Ultra Mac Studio lets its graphics chip use up to 512 GB of shared memory, so two hold the whole model. Each token reads only a small slice, which Meta says the chip's 1.2 TB/s memory speed handles. The sources cite no independent performance results. Meta released the weights under its Muse Model License and said the commercial terms are limited, though many developers can use it free. The sources do not list those terms, and the full evaluation report has no date yet.
For buyers, the choice between renting an API and running a model in-house now turns on hardware, licence and data location. Three questions follow: what the licence permits, where data is processed, and what in-house running would cost. VentureBeat noted that where a model is built does not decide where data is processed.