Meta’s free Muse and Meta AI now run on models taught by Wonder

Meta is rolling out specialized versions of its existing models, trained using its frontier Muse Wonder model, to people who use Muse and Meta AI. They carry two Wonder features, Active Memory and Active Focus, to everyday users.

According to Meta’s research blog, dated November 6, 2026, the new models are meant to give users some of Wonder’s capabilities for complex work and daily tasks. Meta AI chief Alexandr Wang and Meta’s AI account announced the change on X the same day.

Memory that tells temporary from permanent

Active Memory was introduced with Muse Wonder. Meta says the models were trained to separate short-lived information from lasting facts, reason across time, spot stale details, and recognize when something said earlier becomes relevant again.

Active Focus is about persistence. Meta says it “enables Muse to continue working toward a goal, even when faced with unexpected challenges.” The models were trained on real-world bottlenecks so they can try harder, recover or find alternatives. Meta adds that they handle several tasks at once, know when to work alone or collaborate live, and adapt when a user changes course mid-task.

For Meta AI, the company says the same work makes it a better thinking partner: it breaks complex problems into steps, asks sharper follow-up questions and copes with mid-conversation changes of direction.

The numbers Meta is claiming

Meta reports these figures, all from its own evaluations:

• Its Personal Superintelligence Index score for the current Muse model is 47%, up from 38% for the previous one.
• On an Active Memory evaluation, the share of tasks that correctly use critical context reaches 90%.
• Active Focus brings a 67% improvement in completing everyday tasks where unexpected issues arise.

None of this has been independently verified, and Meta does not publish the test sets in the post.

What is still unclear

The post does not name which existing models were retrained, and gives no regional rollout schedule beyond saying it is reaching Muse and Meta AI users. It also does not say whether the specialized models match Wonder itself. The New Stack’s Frederic Lardinois, as summarized by Techmeme, reported that Meta is keeping Wonder for harder tasks for subscribers.

That fits how Meta has split the family. Muse Wonder launched in October 2026, available through the Meta Model API and Muse Code, while Muse Wonder Full is limited to Muse Code subscribers, according to Meta’s launch material. Ordinary users therefore get distilled behavior, not the frontier model.

A distribution play against bigger rivals

The competitive logic is about reach. Runtimewire reports that Muse has passed 10 million setups in North America and frames the update as a test of whether better memory and task recovery can turn Muse into a daily habit, set against ChatGPT, Claude and Gemini. Bloomberg ran a piece on how Meta’s Muse and OpenAI’s Dots agents work.

Meta has been adding features quickly. It added voice mode and expressive avatars to the Muse app in the US, as Techmeme and The Verge reported, and Muse can now build trackers, planners and games from a chat message. Memory and persistence are what would make those features stick, since an assistant that forgets a goal between sessions undermines them.

The approach, using a large model to train smaller ones that serve the mass market, lets Meta ship quality gains without the serving cost of a frontier model for every user. Meta has not shared cost figures, so that remains an inference from its product split, not a stated fact.