Meta released Muse Wonder and Muse Wonder Full, two models the company calls its most capable to date and tunes for agents that carry out personal and professional tasks. Muse Wonder is open to developers now through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers.
Muse Wonder builds on Muse Spark, the model family Meta introduced in April, and was trained at the gigawatt-scale Prometheus data center, Meta says. The company describes the pair as a step on its "scaling ladder" toward what it calls personal superintelligence.
Two behaviors anchor the design. Active Memory carries relevant details from one conversation into the next. Active Focus keeps an agent working when a task turns difficult: it tries other approaches, recovers from errors, asks for clarification and checks in before doing anything important. The target jobs range from untangling an insurance claim or a lease to building a working app, with Meta pointing to domain work in law, finance and health alongside coding.
Meta claims leading results for personal and professional agents, memory and coding, citing Harvey's Legal Agent Benchmark, FocusBench and SWE-Marathon. The numbers sit in charts inside Meta's research post and its accompanying report, and they are Meta's own comparisons against other frontier models, so no outside lab has confirmed them. Meta also argues that Muse Wonder Full handles tasks it was not tuned for better than the trend from Muse Spark 1.3 would predict, crediting larger base models.
A personal agent can reach messages, calendars and private documents, which is why safety gets its own section. Meta says it trained the models to withhold private information without permission and to turn down requests to help with harmful acts. It lists its strongest safeguards yet against chemical, biological, radiological, nuclear and cyber misuse, while keeping defensive cyber work available under its Meta Superintelligence Scaling Framework. A fuller safety report is promised with no date attached.
The rollout order matters here. Developers and Muse Code subscribers get the models first. The consumer Muse app follows soon, subscribers first, Meta says, without giving a date. Until then, the people positioned to test Muse Wonder against Meta's claims are the ones writing code with it.
