Meta’s New Muse Models Hit 47% on Its Own Index, but They Have No Name, No Size and No Training Details

Meta said on 6 November that Muse and Meta AI now run new Muse models trained with Muse Wonder, the frontier model it released on 12 October. Meta calls them specialized versions of its existing models, with two Wonder features: Active Memory and Active Focus.

The evidence is three numbers. Meta’s Personal Superintelligence Index rises from 38% to 47%, its Active Memory evaluation rises from 76% to 90%, and models in Muse show a 2.4 times reduction in forgetting critical information. Meta also claims a 67% improvement in finishing everyday tasks when something unexpected goes wrong.

The catch is everything Meta did not publish. The new models have no names, no parameter counts, no context windows and no model card. Meta did not say how it trained them beyond the phrase “trained using Muse Wonder,” and it gave no baseline for the 67% figure.

Quick answer: Meta has replaced the models behind Muse and Meta AI with new specialized models trained using Muse Wonder. Meta reports a 9 point gain on its own Personal Superintelligence Index, from 38% to 47%, and a 14 point gain on its Active Memory evaluation, from 76% to 90%. The models are unnamed, their sizes are undisclosed, and there is no public model card or independent benchmark. Muse Wonder remains the paid API model at $4 per million input tokens and $20 per million output tokens. The upgrade is real for consumers, but it is not a new frontier release.

What Meta actually shipped

Meta framed the release as a distribution move, not a scaling milestone. The specialized models carry some of Wonder’s behavior to the much larger audience on Muse and Meta AI.

Active Memory is the more concrete feature. Meta says the models now distinguish temporary details from permanent ones, reason across time, and decide which earlier facts are still relevant. Active Focus keeps an agent working when a booking site breaks or a form changes, and lets it take a mid-task correction without losing the thread.

Both features arrived with Muse Wonder in October. The new part is that they now reach users who are not paying for Wonder. Subscribers still get Wonder itself for harder tasks, according to The New Stack, while Wonder Full remains limited to the two most expensive Muse Code plans.

What Meta did not ship is a way to inspect the thing. There is no model name, no API endpoint and no licence. Developers cannot call these models directly.

The 47% number belongs to Meta

The headline is a single chart: 38% for the previous Muse model, 47% for the current one, on a metric Meta calls the Personal Superintelligence Index. The release does not define its tasks, weighting, sample size or confidence interval.

That does not make the number false. It makes it unauditable. A nine point jump is large enough to matter, and the direction is probably real. But a company index cannot tell a buyer how the model compares with Claude, Gemini or GPT on the same work.

Active Memory has the same problem. Moving from 76% to 90% on an internal evaluation is meaningful if it resembles real use. Meta has not released the test set, so the only check is whether Muse remembers a detail from last week.

Metric	Previous Muse model	New specialized models	Change	Source
Personal Superintelligence Index	38%	47%	+9 points	Meta, 6 November 2026
Active Memory evaluation	76%	90%	+14 points	Meta, 6 November 2026
Forgetting critical information	Baseline	2.4 times less	Not quantified	Meta, 6 November 2026
Completing tasks after unexpected issues	Baseline	67% better	Baseline not published	Meta, 6 November 2026
API price for Muse Wonder	Not applicable	$4 input, $20 output per 1M tokens	Specialized models not sold	OpenRouter, 7 November 2026

Source: Meta research blog and OpenRouter model listing. Missing baselines reflect what the sources did not publish.

A distillation story with the method removed

Training smaller models with a stronger one is standard practice. It is how labs put frontier behavior into consumer products without serving a frontier model to every free user. Meta says it trained the new models on behaviors that matter in a personal agent, such as when to act alone and when to ask.

That is sensible product strategy and the most interesting technical claim in the release, yet Meta stopped at one sentence. There is no description of the teacher-student setup, data mix, compute, student architecture or transfer evaluation.

The silence matters because consumer agents handle private context: messages, calendars, bookings and documents. A model that remembers more is useful, but also a bigger store of personal information. Meta has not published how the models decide what to retain or how a user can inspect it.

Today we're bringing specialized models trained using Muse Wonder to @Muse and Meta AI, giving users access to some of Muse Wonder's frontier capabilities for complex tasks at work and in everyday life.

These specialized models include Muse Wonder's Active Memory and Active… pic.twitter.com/XCjnZMoXaf

— AI at Meta (@AIatMeta) November 6, 2026

Why the naming gap is not cosmetic

A model name is the handle that lets outsiders compare, cite and test a system. Without one, a benchmark result cannot be tied to a version, a price or a licence. Meta’s developer site lists Muse Wonder, Muse Spark, Muse Glimmer and Muse Image as named products. The new consumer models are not on that list.

The New Stack reported that Meta has not named the models or detailed its method, consistent with the release. Meta also gave no baseline for the 67% figure, so the gain could be from a low base or a high one.

This is a familiar pattern: ship the capability, withhold the specification. It is less defensible when the same company is asking regulators and publishers to trust its AI terms, a debate this blog covered in the UK competition probe into Meta’s commercial terms.

What it means for Indian developers

Indian developers cannot use these models. They are consumer features inside Muse, WhatsApp, Instagram, Facebook and Messenger, with no API, no model ID and no published price.

Muse Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens, the same pass-through price OpenRouter lists. One million tokens in and one million out costs $24. An agent that resends a long system prompt every turn will spend faster than the headline suggests.

Meta’s Muse Spark Open release put an 816 billion parameter model under open weights, though local use needs about a terabyte of memory.

Data residency deserves a flag. The release says nothing about where the specialized models run or what memory they retain. Teams in health, finance or government work should assume prompts and context leave India until Meta publishes a region.

Test memory behavior on your own language and workflow, including Hindi and other Indic inputs, because Meta published no multilingual results for these models. A 90% score on an internal English-heavy evaluation is not a guarantee for an Indian support agent.

Confirmed versus not confirmed

Claim	Status	Evidence
Meta rolled out new specialized models to Muse and Meta AI on 6 November 2026	Confirmed	Meta research blog
The models were trained using Muse Wonder	Confirmed	Meta research blog and The New Stack
The Personal Superintelligence Index rose from 38% to 47%	Confirmed as a Meta figure	Meta chart in the release
The Active Memory evaluation rose from 76% to 90%	Confirmed as a Meta figure	Meta chart in the release
Task completion improved 67% when unexpected issues arise	Partly confirmed	Meta states the gain but publishes no baseline
The new models beat Gemini, Claude or GPT on agent tasks	Not confirmed	No independent benchmark or cross-vendor comparison
The models have names, published sizes or a model card	Not confirmed	Meta published none; The New Stack says they are unnamed
The training method and data are disclosed	Not confirmed	Meta describes the approach in one sentence
The specialized models are available through an API	Not confirmed	No API listing or price exists for them

FAQ

What did Meta announce on 6 November 2026?

Meta rolled out new specialized models to Muse and Meta AI users. It says they were trained using Muse Wonder and bring Active Memory and Active Focus to consumers. The models have no public names.

Is this a new frontier model?

No. Muse Wonder, released on 12 October 2026, remains the frontier model. The new release applies its behavior to smaller models. Subscribers still get Wonder for harder tasks.

How much better are the new models?

On Meta’s own numbers, the Personal Superintelligence Index rises 9 points to 47%, the Active Memory evaluation rises 14 points to 90%, and forgetting of critical information falls 2.4 times. None of these figures has been independently verified.

Can developers use these models?

Not directly. They are consumer features with no API, model ID, price or licence. Developers can use Muse Wonder through the Meta Model API at $4 per million input tokens and $20 per million output tokens.

Why does the missing model name matter?

Without a name, version and model card, outsiders cannot reproduce the results, compare the model with rivals or track changes. The numbers remain vendor claims, not testable specifications.

Sources

Meta AI Research, Bringing Specialized Models to Muse and Meta AI, 6 November 2026.

Meta AI Research, Muse Wonder: A Frontier Model For Personal and Professional AI Agents, 12 October 2026.

The New Stack, Meta’s best AI model now trains the one most of its users get, 6 November 2026.

OpenRouter, Meta: Muse Wonder model listing and pricing, accessed 7 November 2026.

Earlier coverage: the UK CMA probe into Meta’s commercial terms, Muse Spark Open’s 816B weights, and Thinking Machines’ benchmark appendix.