Meta releases 816B-parameter Muse Spark Open for coding and agent work

Meta has put open weights behind Muse Spark, the company’s large AI model for coding and long-running agent tasks. Until now it had been proprietary. The new release is called Muse Spark Open, and Meta is pitching it as the strongest open-weight model yet from a U.S. lab.

The size is the headline number: 816 billion parameters, though only 42 billion activate for each token. It can take in text, images, audio and video in the same prompt, and Meta says the context window stretches to 1 million tokens. That is a lot of room for long jobs, which is exactly where Meta wants this model to work.

This version starts from Muse Spark 1.2, which shipped in August, and Meta added safety post-training so it can function without the company’s own service guardrails. Meta says the model cleared the release thresholds in its Superintelligence Scaling Framework, with a full evaluation report still pending. Hugging Face’s technical writeup says the training leaned on “real developer workflows,” including multi-file refactors and long debugging sessions.

Meta also tested Muse Spark Open outside its own data centers. It ran on two Apple Mac Studio desktops with M5 Ultra chips linked by Apple’s MLX framework, with about a terabyte of memory between them. On that setup, Muse Code handled bug fixes and new features for the open-source Mastodon client Ice Cubes. Another test used DaVinci Resolve, where the model went through footage frame by frame to find and crop shots for vertical video.

The benchmark picture is mixed, and that matters. Meta says Muse Spark Open scored 61.8 on JobBench, more than twice the 31.1 from Thinking Machines Lab’s Inkling, and ahead of Nvidia’s Nemotron 3 Ultra at 27.2. But Chinese open models still lead on coding: Kimi K3 hit 85 on Terminal-Bench 2.1, while GLM 5.3 and Qwen3.8 both beat Muse Spark Open’s 76.4. On tool use, though, Meta says its model tops the list. Its 89.2 on MCP-Atlas beat every rival Meta named, including the Chinese ones.

The release comes with a new Muse Model License that allows developers to adapt the model or ship it in their own software without a fee. Companies that sell it as a hosted service and have crossed $20 million in gross revenue over 12 months need a separate agreement, and products built on it must carry a “Built with Muse Spark Open” label. Military and critical infrastructure uses are banned. Meta’s previous open-weight model was the much smaller Muse Glimmer in August, while its most capable systems are still closed. The new Muse Wonder models introduced Monday are only available through the Meta Model API and Muse Code, though Muse Spark Open itself is now on Hugging Face and in the model catalogs of Amazon Web Services, Google and Microsoft, with seven inference providers offering hosted access at launch.