Meta releases Muse Spark Open, an 816-billion-parameter open-weight model tuned for Macs

Meta claims its most capable US open-weight model yet

Meta has released a large AI model called Muse Spark Open, and it is making a striking claim about where the model stands: according to Meta, as reported by The Verge, it is the "most capable US open-weight model to date." The weights were published on the day of the announcement, meaning the model can be downloaded and self-hosted rather than accessed only through an API.

The model has 816 billion total parameters and is, in Meta's words, a "natively multimodal model." The company says it performs well on long-horizon tasks, complex software engineering, and multi-step reasoning.

Built to run on two Mac Studios

The more unexpected part of the launch is the hardware story. Meta says Muse Spark Open is the first model in its class that can run on a desktop — specifically, on two Apple M5 Ultra Mac Studios using the latest MLX framework. MLX is Apple's array framework for machine learning tasks on Apple Silicon.

Apple participated directly in the launch. The Verge reports that Apple published a new technical post covering MLX, Muse Spark Open, and two other open models.

An unusual public alliance

Meta and Apple are normally rivals, but the two companies exchanged public endorsements for the release — and according to Axios, they had been collaborating for several months before the announcement.

Meta CEO Mark Zuckerberg wrote on X that the model, with weights releasing the same day, "runs great on new Mac Studios thanks to work with" Apple's John Ternus. Ternus, whom The Verge identifies as Apple's new CEO, replied with congratulations, saying it is remarkable to run a model of that caliber and scale efficiently right from a desk and that Mac Studio "keeps pushing the boundaries of on-device AI."

Meta chief AI officer Alexandr Wang was blunter in a remark to Axios about the two companies cooperating: "We're excited for mommy and daddy to be playing nice again."

Why it matters

Three things make this release notable. First, the open-weight claim: if independent evaluations support Meta's positioning, Muse Spark Open would push the frontier of freely downloadable models from US labs, with everything that implies for researchers, startups, and anyone who wants to self-host capable AI rather than rent it.

Second, the hardware milestone. A model with more than 800 billion parameters running on a pair of desktop machines — not a rack of datacenter GPUs — reframes what local inference can look like. That is partly a software achievement through Apple's MLX framework, and it hands Apple Silicon a concrete high-end AI credential just as on-device and on-desk AI becomes a competitive battleground.

Third, the partnership itself. Meta and Apple have often been at odds, and a months-long engineering collaboration capped by mutual public praise marks a shift in tone rather than a one-off press exchange. For Meta, Mac optimization extends the model's reach; for Apple, it supplies a flagship open model that shows off its hardware.

The caveats are worth keeping in view: the capability claims are Meta's own, "first in its class" is doing a lot of work in that sentence, and two M5 Ultra Mac Studios are professional-grade machines, not consumer desktops. Even so, as a signal of where open-weight models and local inference are heading, the launch is a meaningful data point.