Meta has released the weights of Muse Spark Open, a model with 816 billion parameters that it says runs on two Apple Mac Studio desktops. Models of that size normally live in data centres. Meta’s pitch is that this one does not have to.

The trick is arithmetic. Of those 816 billion parameters (the numbers a model learns in training), only 42 billion are used for any single word it processes. A model built this way stores a great deal but computes a little, so the expensive part is holding it in memory, not running it. Meta’s post says each Mac Studio with the M5 Ultra chip can give its graphics processor up to 512 GB of memory, and that the chip moves data at 1.2 TB per second, fast enough to fetch the small slice each word needs. Two machines together offer about a terabyte.

Meta’s research blog, which published the announcement on 15 October, calls the setup “the first model in its class to run on your desktop.” It does not say what precision the weights are stored in, and that matters. Our arithmetic: at standard 16-bit precision, 816 billion parameters would need roughly 1.6 TB, more than the pair offers, so the setup would have to rely on a compressed version of the model. The post also gives no tokens-per-second figure and no price. Two top-spec Mac Studios are not a laptop purchase, and nothing in the post suggests otherwise.

What the model does is described in general terms. Meta calls it natively multimodal, meaning it was built to handle more than text, but the post never lists which kinds of input it accepts. It holds a context of one million tokens, roughly the amount of text it can keep in view at once, and Meta says it is built for coding and long, multi-step tasks. For proof, the company says it used the model through its own coding tool, Muse Code, running entirely on the desktops, to ship bug fixes and new features to an open-source Mastodon app called Ice Cubes, and to drive the video editor DaVinci Resolve, finding and cropping the strongest moments in footage.

The headline claim is that this is “the most capable US open-weight model to date.” That is Meta’s description of its own release. The post has a section titled Key Benchmarks, but the text contains no figures and names no rival models, and the evaluation report is linked rather than reproduced. A fuller safety evaluation, which Meta says met the thresholds in its own scaling framework, is promised “soon.” Nothing here is independently measured.

The licence deserves the most caution. Meta says the weights come under the Muse Model License “with limited commercial terms, which allow many developers to use the model for free.” That is all the post says. It does not state what the limits are, who falls outside “many developers,” or whether there is a usage or revenue cap. A licence like that is not open source in the usual sense, and the word “open” in the model’s name tells you about the download, not the rights. The post also does not say where this model sits relative to Meta’s other models, so whether it is a flagship or a step below one is unanswered.

Distribution is wide. Hosted versions run through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, and weights are downloadable from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.

Anyone planning to build a product on Muse Spark Open should read the Muse Model License before the benchmarks, because the commercial terms decide whether the model can ship, and Meta has so far described them in one sentence.

Based on Meta’s research blog post “Introducing Muse Spark Open” of 15 October 2026.