THE ESSENTIALS IN 30 SECONDS

  • On November 6, Meta said it is rolling out specialized versions of its existing models, trained using Muse Wonder, to people who use Muse and Meta AI.
  • The update centres on Active Memory, for recalling context that matters, and Active Focus, for pushing through obstacles during a task.
  • Meta reports a Personal Superintelligence Index of 47% versus 38% for its previous Muse model, but publishes little detail on how the benchmarks work.

Meta Superintelligence Labs said on November 6 that it is rolling out specialized versions of its existing models to people who use Muse, its personal agent app, and the Meta AI assistant. According to the company's research blog, these versions were trained using Muse Wonder, the frontier model family Meta introduced on October 12. The post does not say which existing models were specialized, nor does it give dates, regions or subscription tiers for the rollout.

What changes for users

The update focuses on two capabilities Meta first described with Muse Wonder. Active Memory is meant to help the assistants keep track of what matters and notice when something said earlier becomes relevant again. Meta says the models were trained to distinguish temporary from lasting information, to reason across time and to set aside details that have gone stale.

Active Focus is designed to keep Muse working toward a goal when a task hits an unexpected obstacle. Meta says it trained the models on real-world bottlenecks so they can retry, recover or look for another route, and that they can better judge when to continue alone and when to check in with the user, as well as absorb corrections mid-task. For Meta AI, the company describes the changes in qualitative terms: breaking problems into steps, asking sharper follow-up questions and adjusting when a user changes direction.

The numbers Meta reports

All figures come from Meta and compare the new Muse model with the previous one. On what Meta calls its Personal Superintelligence Index, the score rises to 47% from 38%. On an internal Active Memory evaluation, the share of tasks where critical context is used correctly goes from 76% to 90%. Meta also claims a 2.4x reduction in forgetting critical information needed to reach users' goals and a 67% improvement on everyday tasks where unexpected issues come up.

The blog post does not explain how these benchmarks are built, what tasks they contain or what baselines underlie the 2.4x and 67% figures, and it includes no comparisons with competing assistants. None of the results has been independently verified.

Context

When Muse Wonder launched in October, it was offered through the Meta Model API and Muse Code, and Android Headlines and Digit reported that access in the Muse app was tied to paid subscriptions. The new rollout appears to extend Wonder-derived capabilities to a wider base of Muse and Meta AI users, though Meta has not detailed who gets them first. Muse has also recently gained voice conversations, avatars and Spaces, which Meta announced separately.