On 15 October Meta released Muse Spark Open, in its own words the most capable open-weight model from the US. Of its 816 billion parameters only 42 billion work on any given token, because the model routes each request to a few specialised sub-networks. It reads images as well as text, takes up to a million tokens of context and is built for coding and agents. The weights are on Hugging Face and at AWS, Google Cloud and Microsoft; Together AI, Groq and OpenRouter, among others, offer it as a hosted service.
Open weights means Meta hands out the finished, trained model, but not the training data or training code, and the model comes under Meta's own licence rather than an open-source one. Using it internally, for research or inside your own product needs no permission. Offering it to others as a service is different: once your group turns over more than 20 million dollars a year, you have to notify Meta first and sign a separate agreement. On top of that come a mandatory “Built with Muse Spark Open” notice in the interface and a list of prohibited uses. On Hacker News, much of the discussion was about whether “open” is the right word at all.
For companies, the point is that a model of this size can now run in their own data centre. The data never leaves the building, and nobody can retire the model or raise its price. That is not cheap: for a desktop setup Meta names two Mac Studios with 512 gigabytes of memory each, which commenters put at around 30,000 dollars. Nor does Muse Spark Open reach the top. In Meta's own table it scores 39 on the Artificial Analysis Intelligence Index, while the Chinese models GLM 5.3 and Kimi K3 reach 45 and 44. If you need an open model from the US, this is now the strongest one; if you simply want the strongest open model, it is still somewhere else.
