Meta Platforms Inc. released open weights today for Muse Spark, the artificial intelligence model it had kept proprietary since launching in April. The open release, called Muse Spark Open, targets coding and long-running agent work, and Meta says it is the most capable open-weight model to come out of a U.S. lab to date. For any single token, only 42 billion of the model's 816 billion parameters are active. Prompts can combine text with images, audio or video, and the context window reaches up to 1 million tokens.
Meta built the release on Muse Spark 1.2, which shipped in August, then added safety post-training so the model performs without the protections of Meta's own services. Meta said Muse Spark Open met the release thresholds in its Meta Superintelligence Scaling Framework, and a full evaluation report is still to come. A technical writeup from Hugging Face Inc. says training drew on what Meta calls real developer workflows, including multi-file refactors and long debugging sessions.
Away from the data center, Meta tested the model on two Apple Inc. Mac Studio desktops with M5 Ultra chips joined into one system by Apple's MLX framework. The machines held roughly a terabyte of memory combined. Muse Code ran on that setup while the coding terminal worked through bug fixes and new features for Ice Cubes, an open-source Mastodon client. The model also processed footage in DaVinci Resolve frame by frame, selecting and cropping the best shots for vertical video.
In Meta's benchmark table, Muse Spark Open scored 61.8 on the JobBench agent test, more than double the 31.1 posted by Inkling, the U.S. open model from Thinking Machines Lab Inc. Nvidia Corp.'s Nemotron 3 Ultra sat lower at 27.2. Chinese open models remain ahead on coding, with Kimi K3 topping the Terminal-Bench 2.1 test at 85, while GLM 5.3 and Qwen3.8 both exceeded Muse Spark Open's 76.4.
Tool use reverses that picture. Muse Spark Open's 89.2 on MCP-Atlas beat every rival Meta listed, including the Chinese models. That benchmark gives a model plain-language tasks without naming the tools required, leaving the model to determine which of 220 tools across 36 live Model Context Protocol servers to call and in what order.
The weights carry a new Muse Model License with what Meta calls limited commercial terms, and developers who adapt the model or ship it inside their own software pay no fee. Meta requires a separate agreement from any company that sells the model as a hosted service and has passed $20 million in gross revenue over a 12-month period. Products built on it must display a Built with Muse Spark Open label, and military and critical infrastructure uses are prohibited.
Meta's prior open-weight release was the far smaller Muse Glimmer, a 30 billion-parameter model that arrived in August. Its strongest systems stay closed, and the Muse Wonder models introduced Monday are offered only through the Meta Model API and Muse Code. Seven inference providers supply hosted access at launch, among them Groq Inc. and Fireworks AI Inc., and the weights can be downloaded from Hugging Face or the model catalogs of Amazon Web Services Inc., Google LLC and Microsoft Corp.
More AI news from TechManNews.
