Summary: On October 12, 2026, Meta introduced Muse Wonder and Muse Wonder Full, its most capable models to date, aimed at agentic tasks in law, finance, health, and coding. The interesting part for builders is not the benchmark language but the two named capabilities: Active Memory and Active Focus. Both are available today through the Meta Model API and Muse Code.
The problem the models are shaped around
Per the Meta announcement, the design targets real-world work: navigating an insurance claim, understanding a lease, making sense of someone’s finances. Those tasks share a painful property — they span multiple documents, stretch across days, and fail quietly when the model loses the thread. Most agent demos break exactly there.
Meta’s answer is two first-class features. Active Memory carries relevant details across conversations and brings them back when the task needs them. Active Focus keeps the model working when a task gets hard: trying other approaches, recovering from errors, and checking in before acting on anything important. I looked at how this framing treats memory and focus as product surfaces rather than prompt tricks in an earlier post on Meta’s agent model memory and focus — this release reads as that thesis shipped at frontier scale.
What’s actually available to build on
Muse Wonder is live today via the Meta Model API and Muse Code. The announcement says the models will soon power more of Muse, rolling out to subscribers first. For API builders, the near-term question is practical: how Active Memory state is scoped, priced, and reset. The supplied announcement does not specify those details, and the full safety evaluation report is also still to come.
Safety posture, briefly
Because a personal agent touches messages, calendars, and documents, Meta says the family is trained to avoid sharing private information without permission and to decline requests that assist genuinely harmful acts. Security teams can still use the models to find flaws in their own code. Releases follow Meta’s Superintelligence Scaling Framework, with a fuller safety report promised.
How to read this release
The concrete takeaway: agent capability is moving from single-shot reasoning toward persistence — memory that survives sessions and focus that survives errors. If you’re routing work across models, Muse Wonder is worth an eval on your longest, most interruption-prone workflows, since that’s explicitly where it’s aimed. Keep the caveat in mind that independent performance and safety evaluations aren’t published yet.
Sources
View the original post on X