Meta Muse Spark Open ships 816B open weights

Meta released Muse Spark Open on October 15, 2026, a version of its Muse Spark model whose weights anyone can download. The model has 816 billion parameters, uses 42 billion of them for each token, and reads up to 1 million tokens at once. It matters because Meta's recent Muse models, including Monday's Muse Wonder, were closed, and developers can now run one on their own hardware.

What Meta released

Meta's announcement calls the model "the open-weight version of Muse Spark." Open weights means the trained model file itself is published, so a company can run it on its own servers instead of only renting it through an API.

Muse Spark Open is a mixture-of-experts model. That design splits the model into many smaller expert networks and only switches on a few of them for each token, the word piece a model reads or writes. So although the full model has 816 billion parameters, only 42 billion do work at any one step. That keeps it cheaper to run than its size suggests.

Spec	Muse Spark Open
Total parameters	816 billion
Active parameters per token	42 billion
Context window	1 million tokens
License	Muse Model License
Hugging Face model	meta-models/Muse-Spark-Open-816B-A42B

The context window is how much text the model can take in at once. One million tokens is enough for a large codebase or several long documents. RuntimeWire adds that the model is natively multimodal, meaning it was built to handle more than text.

Meta's page calls it "the most capable US open-weight model to date." RuntimeWire quotes the same line from Alexandr Wang, who leads Meta's AI effort, and notes that the ranking is Meta's claim, not an independent result. Meta's post links to an evaluation report but gives no benchmark scores in the text.

Where to get it and what it runs on

Meta lists four places to download the weights: Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models. For teams that would rather call an API, Meta names seven hosting partners: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. None of the sources gives a price.

Running it yourself takes serious hardware. Meta says the model runs on two Apple M5 Ultra Mac Studios with up to 512 GB of unified memory each, using Apple's MLX framework. RuntimeWire puts the alternative as "substantial GPU memory."

Meta also wrote custom GPU code, called Helion kernels, for the SGLang serving engine. It reports kernel speedups of 1.03 to 3.95 times on Nvidia and AMD graphics cards, and a 3% end-to-end speedup on AMD's MI350.

The license

The weights ship under the Muse Model License, not a standard open-source license. Meta says its commercial terms are limited but let many developers use the model free of charge. The sources do not publish the full terms, so read them before building a product on the model.

A promise kept, late

Meta had said it would open Muse Spark's weights before. When Meta launched its closed Muse Wonder frontier model on October 13, we noted that the open Muse Spark weights promised for September 2 had still not shipped.

The Muse Spark line has moved fast. In August, The Decoder reported that Meta called Muse Spark 1.2 "primarily a coding upgrade." Its standard price was $1.25 per million input tokens and $4.25 per million output tokens.

What this means for developers

You can now test one of Meta's top models without sending data to Meta. That matters for teams with private code or regulated data, which often cannot use a hosted API at all.

Check the hardware math first. An 816-billion-parameter model needs hundreds of gigabytes of memory even with only 42 billion active parameters, so most teams will start with one of the seven hosting partners. Compare their prices before you commit, since Meta did not publish one.

Read the Muse Model License with your legal team. "Limited commercial terms" can mean revenue caps or user limits, and the sources do not say which applies.

Run your own evaluation. Meta has not printed benchmark numbers in its announcement, so test the model on your real coding and agent tasks against the model you use today.