Meta has introduced Muse Wonder and Muse Wonder Full, a pair of models it describes as its most capable systems so far for personal and professional agent tasks. The announcement moves Meta’s Muse family beyond short question-and-answer interactions toward work that can span multiple steps, such as interpreting a lease, navigating an insurance claim, or building an application. 1
Memory and persistence are the pitch
Meta says the models add two features aimed at tasks that unfold over time. Active Memory is designed to carry useful details across conversations and bring them back when relevant. Active Focus is intended to keep pursuing a goal when an approach fails, try alternatives, recover from errors, and ask for clarification or check in before an important action. Meta positions the models for areas including coding, law, finance, and health. 1
Those capabilities are company descriptions, not a substitute for independent evaluation. Meta’s announcement says a fuller report with additional safety evaluations will be published later. It also describes safeguards intended to prevent unauthorized sharing of private information and to refuse assistance with harmful acts. For an agent expected to work across messages, calendars, and documents, how permissions and confirmations behave in real deployments will matter as much as benchmark performance. 1
Where users can access them
Meta says Muse Wonder is available to developers and builders through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers. The company says Muse Wonder will soon power experiences in the Muse personal assistant, with a subscriber-first rollout. 1
Reuters frames the release as part of Meta’s effort to compete more directly with OpenAI and Anthropic after a major investment in rebuilding its AI effort. It also reports that Meta sees subscriptions around Muse as one possible way to turn its AI spending into revenue. 2
What to watch
The practical question is whether the models can reliably complete long tasks while keeping users in control. Meta’s Active Focus description explicitly includes asking for clarification and checking before following through on important actions; that distinction is significant for agents that may handle personal or work data. The coming safety report and broader Muse rollout should provide more evidence about how those safeguards work in practice.
For now, the release gives developers a new Meta model family to test and puts more attention on agent behavior—memory, recovery, and permission boundaries—rather than only on one-shot benchmark scores.