Meta releases Muse Spark Open, an 816B open-weight model

Meta released the weights of Muse Spark Open on Oct. 15, 2026, a language model anyone can download from Hugging Face. Weights are the parameters a model learned during training. With them, anyone with the hardware can run it on their own machines without sending data to an outside service. Meta says it will also reach the cloud platforms of Amazon (SageMaker AI), Google (Model Garden) and Microsoft (Foundry Models).

The model has 816 billion parameters (816B), but only 42 billion work on each token, the unit of text it processes. It is a mixture-of-experts design, in which only part of the network is active at each step. It accepts up to one million tokens of context, takes text, images, video and audio as input, and answers only in text, according to the model card.

Running it at home takes a lot of memory. The card asks for at least 700 GB of VRAM with the compressed NVFP4 version and 2 TB with BF16. Apple measured the first at about 493 GB and published speed tests on its Mac Studio computers, which the company sells.

Meta describes the model as the most capable of its kind in the United States and touts three results, obtained at its highest reasoning level: 39 on the Artificial Analysis index, 89.0 on GPQA Diamond and 80.4 on OSWorld Verified. These are the company's own results, with no independent evaluation published. The same card acknowledges that other models, such as Kimi K3 and GLM 5.3, score higher on the overall index.

The license is called the Muse Model License and sets these conditions, according to the card:

• It allows using, modifying, distributing and selling copies of the model.
• It requires displaying the notice "Built with Muse Spark Open".
• A company that sells access to the model through an API and exceeds 20 million dollars in gross revenue over 12 months needs a prior agreement with Meta.

The full text was not reviewed, so it is not confirmed whether it meets the definition of open source. Mark Zuckerberg, Meta's chief executive, said on X that the model "runs great" on the new Mac Studio, according to The Verge.

Two loose ends remain. Meta and Apple say 816B, but the Hugging Face card lists 822B for the file size, and no source explains the difference. There is also no price or availability in Mexico for the Mac Studio with M5 Ultra in what has been published so far.