Meta's Muse Spark Open, the “best American open model,” still trails two Chinese rivals

Two Mac Studios side by side, each with 512 GB of memory. According to Meta, that is enough hardware to run Muse Spark Open without a data center. On October 15, the company released the weights of this model, which it describes as “the most capable American open-weights model to date.” The word “American” matters, as Meta's own results table shows.

Muse Spark Open has 816 billion parameters, 42 billion of which are active at each step (a “mixture of experts” architecture). It accepts up to one million tokens of context. It understands text, images, video and audio, and replies in text. Meta mainly aims it at coding and agents, meaning tasks where the AI chains actions on its own. It shares the same base model as Muse Spark 1.3, and Meta says it added dedicated safety training, since the model will be used without the safeguards of Meta's own services.

What Meta's numbers really say

The model card published on Hugging Face compares the model with Kimi K3 and GLM 5.3, two open models developed in China. On the Artificial Analysis overall index, as reported by Meta, Muse Spark Open scores 39, versus 44 for Kimi K3 and 45 for GLM 5.3. It also trails on terminal-based coding (Terminal-Bench 2.1: 76.4 versus 85.0 and 83.9), on scientific reasoning (GPQA Diamond) and, against Kimi K3, on computer use (OSWorld; no score published for GLM 5.3).

It comes out ahead elsewhere: on tool use via the MCP protocol (MCP-Atlas: 89.2 versus 82.3 and 84.2), on a professional-tasks test (JobBench) and above all on resistance to misuse. On StrongREJECT v2, where a lower score is better, it scores 0.5 versus 66.6 and 62.9. All of these figures are published by Meta itself. No full independent evaluation was available on October 16, and Meta only says a full evaluation report is coming “soon”; for now, only an evaluation methodology document is online.

“Open,” but not free

The weights are distributed under the Muse Model License, a custom license, not a free-software license like Apache 2.0. What it requires:

• prominently displaying the notice “Built with Muse Spark Open” in any product or documentation that uses it, including retrained versions;
• above $20 million in revenue over twelve months, notifying Meta and signing a separate agreement to resell access to the model as an API. Internal use, research and integration into your own product are not affected;
• complying with a usage policy that notably bans creating malware and bypassing safeguards.

For European users, the contract is concluded with Meta Platforms Ireland. The model card mentions no restriction specific to the European Union. Downloading, however, requires an account and acceptance of the license.

What it takes to run it

The model is available right away through APIs from several hosting providers, including Fireworks AI, Groq, Nebius, OpenRouter and Together AI. Its weights are also offered on Amazon SageMaker, Google Cloud and Microsoft Foundry. Running it locally takes serious hardware: at least 700 GB of GPU memory in the compressed NVFP4 format, and 2 TB at its original precision, according to the model card. NVIDIA's Nemotron 3.5 model, which fits on a single graphics card, plays in a different league.

For Meta, which already charges for its closed model Muse Wonder, this is a strong comeback in open weights. It remains to be seen what independent rankings will say in the coming weeks, against the Chinese models and France's Mistral Large 4.