Meta Says Its Specialized Muse Models Hit 47% on Its Personal Superintelligence Index and 90% on Active Memory. Every Number Is Meta's, and There Is No Price Sheet.
Meta AI Research said on November 6 that it is rolling specialized versions of existing models, trained with Muse Wonder, into Muse and Meta AI, bringing two features it calls Active Memory and Active Focus. The chart claims: Personal Superintelligence Index 47% against 38% before, an Active Memory evaluation of 90% against 76%, 2.4 times less forgetting of critical information, and 67% better completion of everyday tasks that hit unexpected problems. The post names no parameters, no architecture, no license, no API price, and there is no Artificial Analysis row. We read the research blog.
By T.W. Ghost
Meta AI Research posted today, November 6, 2026, that it is rolling "specialized versions of existing models," trained with Muse Wonder, into Muse and Meta AI. Two features come with them: Active Memory, which is about carrying critical context across a task, and Active Focus, which is about finishing everyday tasks when something unexpected happens along the way. The post has a chart for each.
The chart numbers: on Meta's Personal Superintelligence Index, the current Muse scores 47% against 38% for the previous version. On an Active Memory evaluation, 90% of tasks correctly used critical context, against 76%. Meta claims 2.4 times less forgetting of critical information and a 67% improvement in completing everyday tasks where unexpected issues arise. What the post does not have is anything a buyer could act on: no parameter count, no architecture, no license, no API price, and no row on Artificial Analysis or any other independent board. We read the research blog. Here is what held up, which is the shape of the announcement rather than its numbers.
What Meta Said
| Specialized Muse models (Meta AI Research, November 6, 2026) | |
|---|---|
| What | Specialized versions of existing models, trained with Muse Wonder, rolling into Muse and Meta AI |
| Features | Active Memory; Active Focus |
| Personal Superintelligence Index | 47% now vs 38% before (vendor chart) |
| Active Memory evaluation | 90% vs 76% of tasks using critical context correctly (vendor chart) |
| Forgetting | 2.4x less forgetting of critical information (vendor) |
| Everyday tasks with unexpected issues | 67% improvement (vendor) |
| Parameters, architecture, license | Not stated |
| API price | Not stated |
| Independent score | None |
"Specialized versions of existing models" is not a model. There is no named base, no card, and no version string. Whatever shipped is a tuning of something Meta already runs, described by its teacher rather than itself.
The index is Meta's. A Personal Superintelligence Index is a benchmark Meta built to measure the thing Meta sells. A nine-point gain on it is a vendor result on a vendor test, and the Active Memory evaluation is the same.
Two products share one pitch. Muse is the agent product; Meta AI is the consumer assistant. The post rolls the same specialized models into both and does not say they behave the same.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Blog post Nov 6, 2026; specialized models trained with Muse Wonder rolling into Muse and Meta AI | Vendor (Meta AI Research) |
| Active Memory and Active Focus | Vendor |
| PSI 47% vs 38%; Active Memory 90% vs 76%; 2.4x; 67% | Vendor (Meta's charts) |
| Parameters, license, price | Absent |
| Artificial Analysis or other independent row | None |
Who Should Use It
Use the post as a teaser, which is what it is: Meta is telling Muse and Meta AI users that memory and task completion got better by its own measures.
Skip budgeting or benchmarking against it until there is a model card, an API price, or an independent row; a 47% on a vendor index cannot be compared with anything outside Meta.
The frontier is elsewhere: the Meta model with independent numbers is still Muse Spark Open, on a scale that has since changed. Our model comparison has the current standings side by side.
Sources
*Not sure whether a vendor's own superintelligence index tells you anything about your tasks? Take the free 2-minute quiz and get matched. Then read about the Meta model that did publish weights and a score: Muse Spark Open.*