On November 6, 2026, Meta announced on its official research blog that specialized models trained on Muse Wonder are rolling out to Muse and Meta AI users. It is the first capability hand-down since the Wonder series launched a few weeks ago: not another new model, but frontier capability brought into the two products people use every day. Meta's headline number is its Personal Superintelligence Index, up from 38% to 47%.
Active Memory: what it remembers, what it forgets
The core of the upgrade is two capabilities that shipped with Wonder. Active Memory lets a model remember what matters, recognize when old information becomes relevant again, and tell temporary information from permanent, reason across time, and judge what has gone stale. Meta's figures: the improved Muse models forget critical information 2.4× less often, and the share of tasks that correctly use critical context in the Active Memory eval rose from 76% to 90%. This is exactly the problem Agent Memory is about: an agent's biggest weakness is often not that it cannot do the job, but that it cannot remember.
Active Focus: holding on when things go wrong
Active Focus lets a model keep working toward a goal when unexpected difficulties appear: it was trained on the kinds of bottlenecks that show up in the real world, so it tries harder, recovers, or finds an alternative on your behalf, and completion of everyday tasks where something unexpected crops up improved by 67%. The specialized models also learned to multitask, knowing when to push on a goal on their own and when to collaborate live with the user, and they can follow a mid-task adjustment without getting confused. On the Meta AI side, the same improvements make it more of a thinking partner: it breaks complex problems into clear steps, asks sharper follow-up questions, and keeps up when you change direction mid-conversation.
For users, the change is gradual: no new product to switch to, no new interface to learn — the same conversations slowly become more retentive and more patient. The limits are worth noting: these are 'specialized models' whose capability comes from Wonder training, but Meta does not say they are Wonder itself; and the scores are Meta's own index and evals, not third-party benchmarks, so the 47% should not be compared directly with other vendors' leaderboards. The rollout is gradual, so different regions and accounts will get the new models at different times.
Meta gave no firm timetable for full rollout. Two things are worth watching next: whether the model really remembers critical information across conversations in practice, and whether it genuinely keeps up — or quietly drifts — when a multi-step task changes direction midway. The official numbers cannot settle either question; only using it will.
