Meta released Muse Wonder and Muse Wonder Full on October 12, 2026, describing them as its “most capable models” and the successors to the Muse Spark line from Meta Superintelligence Labs. Muse Wonder is available to developers through the Meta Model API and Muse Code, and Meta says it will soon power the Muse assistant, rolling out first to subscribers. Muse Wonder Full is available to Muse Code subscribers.
Meta built the series around two capabilities. Active Memory carries relevant information across conversations, and Active Focus keeps the model working toward a goal when a task gets hard, trying other approaches, recovering from errors and checking in before important actions. Meta targets expert work in law, finance and health, along with coding, and names Harvey’s Legal Agent Benchmark, FocusBench and SWE-Marathon among its evaluations. The models were trained at Meta’s gigawatt-scale Prometheus data center.
On safety, Meta calls Muse Wonder its “most aligned model to date,” says it carries its strongest safeguards so far against chemical, biological and cyber misuse while keeping defensive cyber use available, and says it was released under the Meta Superintelligence Scaling Framework with a full safety report to follow.
Why it matters: Muse Wonder is Meta’s bid to put a closed model at the frontier alongside OpenAI, Anthropic and Google, framed around long-running agents rather than chat, which is where the leading labs were competing by late 2026.
What it does not show: the announcement gives no benchmark numbers in text (results are in charts), no pricing and no context window, and the full safety report had not been published at release.