Meta releases Muse Spark Open with 816B parameters for coding and agent work

Meta Platforms Inc. today released open weights for its Muse Spark artificial intelligence model, which has been proprietary since its debut in April.

The version called Muse Spark Open is designed for coding and lengthy agent work. Meta describes it as the most powerful open-weight model to come out of a US lab to date. Only 42 billion of the model’s 816 billion parameters are activated for a given token. Prompts can mix text with images, audio or video, and Meta has set the context window to hold up to 1 million tokens.

The starting point was Muse Spark 1.2, which shipped in August, and then Meta added security after training, allowing the model to exist without the guardrails of Meta’s own services. According to Meta, Muse Spark Open has exceeded the release thresholds in its Meta Superintelligence Scaling Framework, with a full evaluation report pending. A technical report from Hugging Face Inc. said the training relied on what Meta calls “real developer workflows,” including multi-file refactors and long debugging sessions.

Meta also tested the model outside the data center on two Apple Inc. Mac Studio desktops with M5 Ultra chips connected into a system using Apple’s MLX framework. The two machines had a total of about a terabyte of memory. Muse Code ran on this setup while the coding terminal worked through bug fixes and new features for the open source Mastodon client Ice Cubes. The other test was footage in DaVinci Resolve, and the model went through it frame by frame to select and crop the best shots for vertical video.

In Meta’s own benchmark table, Muse Spark Open scored 61.8 on the JobBench agent test, more than double the 31.1 scored by Inkling, the U.S. open model from Thinking Machines Lab Inc. Nvidia Corp.’s Nemotron 3 Ultra was further behind at 27.2.

Chinese open models are better at coding. Kimi K3 topped the Terminal Bench 2.1 coding test with an 85, and GLM 5.3 and Qwen3.8 both finished above Muse Spark Open’s 76.4.

When it comes to tool use, the order is reversed, and Muse Spark Open beats every listed rival meta, including the Chinese models, with a score of 89.2 on MCP-Atlas. The benchmark gives a model tasks in simple language without specifying the tools required to complete them. It is left to the model to figure out which of the 220 tools on 36 Live Model Context Protocol servers should be called and in what order.

The weights come with a new Muse model license that includes what Meta calls “limited commercial terms,” and there are no fees for developers who customize the model or integrate it into their own software. Meta requires a separate agreement from any company that sells the model as a hosted service and has gross sales of more than $20 million over a 12-month period. Products based on this must bear the “Built with Muse Spark Open” label. Military and critical infrastructure uses are taboo.

Meta’s previous open-weight release was the much smaller Muse Glimmer, a 30 billion parameter model that came out in August. The company’s most powerful systems remain closed and the Muse Wonder models launched on Monday are only available through the Meta Model API and Muse Code.

Seven inference providers are offering hosted access at launch, including Groq Inc. and Fireworks AI Inc. The weights can also be found on Hugging Face or in the model catalogs of Amazon Web Services Inc., Google LLC and Microsoft Corp. be downloaded.

Image: Meta