Muse specialized models: check 47% vs 38% and 90% vs 76%, not "personal superintelligence now remembers you"

Meta Superintelligence Labs published a roughly two-minute research post on 6 November 2026 titled "Bringing Specialized Models to Muse and Meta AI." It says the Muse Wonder series introduced a few weeks earlier has been used to train specialized versions of existing models, and those versions are rolling out to people who use Muse and Meta AI.

This is not a release of the strongest model to everyone. It is a hand-down of capability: Active Memory and Active Focus, first shown with Wonder, are being attached to models that already run in consumer products. What to check is the internal numbers in the post, and the difference between those specialized versions and Wonder itself. Do not read "remembers better and sticks to a goal" as proof that personal superintelligence is ready to use.

Hard facts (https://research.meta.ai/blog/specialized-models-for-muse-and-metaai; datePublished 2026-11-06T00:00:00.000Z; bylined Meta Superintelligence Labs; page marked 2 minute read; accessed 2026-11-08, Asia/Shanghai):

• Background: the post says the Muse Wonder series was introduced a few weeks earlier and calls it the lab's most capable to date, and says it delivered what the post calls state-of-the-art performance on personal and professional agent tasks, meaning tasks where the assistant is supposed to act on your behalf at work or in daily life. The news that day is that specialized versions of existing models, trained using Muse Wonder, are rolling out to Muse and Meta AI users.
• Product framing: bring some frontier capabilities to more people, not announce that Wonder is open to every user. The page's JSON-LD description is "Bringing key frontier capabilities to everyone with specialized models for Muse and Meta AI."
• Personal Superintelligence Index: 47% for the current Muse model, 38% for the previous Muse model. The page charts those two scores and does not give item composition, sample size, or scoring rules.
• Active Memory: Muse and Meta AI are meant to remember what matters and notice when earlier information becomes useful again. Training aims include telling temporary information from lasting information, reasoning across time, and telling what is still relevant from what has gone stale.
• Active Memory figures: in Muse, forgetting critical information needed to help you reach your goals is reduced about 2.4 times. On an Active Memory eval, tasks that correctly use critical context are 90% for the current model and 76% for the previous one.
• Active Focus: keep working toward a goal when unexpected obstacles appear. The post says the models were trained on a variety of bottlenecks meant to reflect the real world, so they can try harder, recover, or find another path for the user. In Muse, completing everyday tasks where unexpected issues arise improves by 67%. The post does not give the baseline completion rate behind that relative gain.
• Other behavior notes: the specialized models are described as better at handling several tasks at once, knowing when to push a goal on their own and when to work live with you, and adapting to mid-task corrections without getting confused. For Meta AI, the post gives only qualitative claims: break complex problems into steps, ask sharper follow-up questions, and keep up when you change direction. It does not publish a separate index or eval score for Meta AI.
• Close: the lab says it wants to give more people access to more capable superintelligence.

Internal comparison figures published by Meta

Evidence strength: this is the lab's own short product-research post, with two internal comparison charts. It is not a peer-reviewed paper, and no third party has reproduced the numbers. What you can check now is whether the page still reports 47% versus 38%, 90% versus 76%, 2.4 times, and 67%, and whether the body still describes the rollout as specialized versions rather than Wonder itself.

The numbers have one visible arithmetic relationship. An Active Memory eval rising from 76% to 90% means the share of tasks that did not correctly use critical context fell from 24% to 10%. 24 divided by 10 is 2.4. That lines up with "about a 2.4 times reduction in forgetting critical information." Meta does not say these are the same measurement. Treating them as two separate breakthroughs may double-count one family of internal results. Treating them as independent experiments that already confirm each other also goes beyond the text.

The 67% figure cannot be converted the same way, because the baseline completion rate is not published. If the old completion rate was low, a 67% relative gain can still leave many failures. If the old rate was already high, the same relative figure implies a larger absolute change. A reader cannot tell which case this is.

The Personal Superintelligence Index rising from 38% to 47% is a 9 point gain and remains below half. Without the item list, it shows progress on Meta's own ruler, not an externally checkable claim that personal superintelligence has been reached.

Strongest steelman: instead of shipping the strongest model to every user, Meta is using it to train models that already run inside Muse and Meta AI. That is a coherent product path. What users notice is an assistant that forgets a constraint, treats outdated information as current, or hands the problem back as soon as a task stalls. Separating temporary facts from lasting ones, noticing when a fact has gone stale, and trying another route after a blockage are aimed at those failures. Reporting an eval as the share of tasks that correctly use critical context is also easier to check than saying the system "has memory."

Steelman risk: the post is about a two-minute read and gives no method, item count, control setup, or failure cases. The index and the Active Memory eval are internal benches. The 67% figure has no baseline. "Superintelligence" appears in the product language and in the lab's name, but a 47% score on the lab's own index does not carry what that word means in ordinary speech. Meta AI is said to benefit in the same way, yet it has no separate scores. If later official material adds method notes, sample items, independent benches, or Meta AI figures, raise the score for checkable completeness. Those details are not in this post.

For Muse and Meta AI users, the next six months to watch are whether constraints and preferences are remembered more steadily across conversations, whether stale facts are treated as current less often, and whether the assistant finds another path when a task stalls instead of stopping at once. For eval watchers, the question is whether Meta publishes methods for the Personal Superintelligence Index and the Active Memory eval, and the baseline completion rate behind the 67% figure. For product watchers, check whether access to Wonder itself changes alongside this "specialized versions" post. The text is about those versions entering Muse and Meta AI, not about Wonder opening to everyone.

What to check in this lab post is that specialized versions of existing models, trained using Muse Wonder, are entering Muse and Meta AI, plus the internal figures of 47% versus 38%, 90% versus 76%, about 2.4 times, and 67%. It is not that personal superintelligence now steadily remembers you and finishes the job. The numbers are official. Most of the method and the missing baseline are not. Reading the short post as proof that memory is solved leaves the source behind.

【No.2】