Meta opens Muse Spark's weights: a huge model that (almost) runs at home
Muse Spark Open has 816 billion parameters, uses 42 per word and can run on two Mac Studios. It's the most capable American open model so far, with a licence that isn't fully free.
by Mattia Panatta
On 15 October Meta published the weights of Muse Spark Open, the downloadable version of the model it had kept to its own services since April. Anyone can download it from Hugging Face, modify it and use it in their own software.
The numbers
- 816 billion parameters in total, 42 billion active for each word: it’s a “Mixture of Experts” model that switches on only part of itself at a time.
- A one-million-token context and the ability to read images too.
- According to Meta, it can run locally on two Mac Studios with M5 Ultra chips linked together.
In Meta’s published tests it’s the best at using external tools (the MCP-Atlas benchmark), while in coding it trails Chinese open models such as Kimi K3, GLM-5.3 and Qwen3.8. These are Meta’s figures: I’d wait for independent measurements.
The licence
“Open weights” doesn’t mean “free to use”. The new Muse Model License is free for anyone who adapts the model or ships it inside their own software, but anyone offering it as an online service with more than $20 million in annual revenue must sign a separate agreement with Meta. For a professional or a small business it’s effectively free; for large providers it isn’t.
My take
It’s good news for anyone who cares about data confidentiality: an American open-weights model at this level was missing. But don’t read “runs on two Macs” as “runs on your computer”. Two Mac Studios with the required memory cost many thousands of euros, and with this year’s RAM prices the figure goes higher still. I discuss this in the lesson on local AI: for most people a business plan for a cloud model remains more capable and far cheaper.
For companies whose data can’t leave the premises, though, Muse Spark Open becomes a serious option worth testing.
