Muse and Meta AI Switch to Specialized Wonder-Trained Models
Starting today, Muse and Meta AI run on specialized models trained using Muse Wonder, with Active Memory for recall across conversations and Active Focus for pushing through blocked tasks.
By Icarus2 min read
Muse and Meta AI now run on specialized models trained using Muse Wonder, the frontier model family Meta introduced in October. Alexandr Wang, Meta’s chief AI officer, announced the switch on X, saying the new models started rolling out today.
The models are specialized versions of Meta’s existing models, trained using Muse Wonder rather than Wonder itself. They carry two behaviors Meta calls Active Memory and Active Focus. Active Memory lets the models reason across time and remember what matters from earlier conversations, distinguishing temporary details from permanent ones and current information from stale information. Active Focus keeps work moving toward a goal when a task hits a roadblock, trying other approaches, recovering from errors, and checking in before acting on anything important.
The Numbers Meta Is Reporting
In Muse, the change shows up as a 2.4x reduction in forgetting critical information and a 67% improvement in completing everyday tasks when things don’t go to plan, according to Meta’s research post. On Meta’s Personal Superintelligence Index, the current Muse model scores 47%, up from 38% for the previous model. On Meta’s Active Memory evaluation, the current model used critical context correctly in 90% of tasks, compared with 76% previously. These figures come from Meta’s own evaluations.
| Benchmark | Previous Muse model | Current Muse model |
|---|---|---|
| Personal Superintelligence Index | 38% | 47% |
| Active Memory Eval, tasks correctly using critical context | 76% | 90% |
Data from the two charts in Meta's research post and Alexandr Wang's announcement; the 2.4x and 67% figures above are reported separately in the post text, not in the charts.
The specialized models also handle multitasking, decide when to pursue a goal on their own or work live with the user, and adapt to steering and corrections mid-task. In Meta AI, the same changes break complex problems into steps, ask sharper follow-up questions, and accommodate changes of direction during a conversation.
The rollout brings Wonder-derived models to the consumer apps a few weeks after Muse Wonder and Muse Wonder Full launched for developers and Muse Code subscribers. It lands in the same week as two other Muse releases: Muse Voice, which opened real-time voice and video chats to everyone in the US, and Spaces, the custom experiences Meta introduced today.
