Meta's new frontier models are built for agent work

Meta has released Muse Wonder and Muse Wonder Full, which it calls its most capable models so far and pitches squarely at agents: systems that carry a task through rather than answer a single prompt. The examples Meta chose are deliberately ordinary — working through an insurance claim, reading a lease, making sense of someone's finances, building a working app.

Access is split by tier. Muse Wonder is available now to developers through the Meta Model API and Muse Code, while Muse Wonder Full is limited to Muse Code subscribers. Meta says Muse Wonder will later power its consumer Muse experiences, rolling out to subscribers first.

Two named capabilities carry the release. Active Memory is meant to keep what matters about a user across conversations and bring back the details relevant to the task in hand. Active Focus is about persistence: staying on a goal when a task gets hard, trying other approaches, recovering from errors, and asking for clarification or checking in before doing anything important. Meta's accompanying research post adds that the models were trained at its gigawatt-scale Prometheus data centre and sit on larger base models than Muse Spark, which it credits for better results on tasks that were not targeted during post-training.

Because a personal agent works across messages, calendars and documents, Meta says it trained the family not to share private information without permission and to refuse help with genuinely harmful acts, while keeping the ability to find flaws in code for defenders. The models are released under its Meta Superintelligence Scaling Framework. The full report with additional safety evaluations is not out yet; Meta says it will follow soon.

The launch comes fourteen months into what Meta describes as a ground-up overhaul of its AI work. Its own framing is that the company has moved from rebuilding foundations to competing at the frontier, with Muse Wonder as the first release meant to show it.