Meta gives all Muse and Meta AI users Wonder-trained models, claiming 2.4x less forgetting

What Meta announced

Meta said on 6 November 2026 that it is bringing specialized models to Muse and Meta AI. These are new versions of the existing models behind its two consumer assistants, trained using Muse Wonder, the frontier model Meta launched in October. Mark Zuckerberg wrote on X that the update brings two of Wonder's capabilities to everyone, and coverage of the post confirms the rollout includes users who do not pay.

The two capabilities are the ones Meta introduced with Muse Wonder. Active Memory is about keeping and retrieving the right details across a conversation and across sessions. Active Focus is about persistence: the assistant keeps working toward a goal, tries a different approach when one fails and adapts when new information appears. Muse Wonder itself is not part of this rollout and remains the tier for harder tasks, available to subscribers and to developers.

Meta describes the rollout as automatic: users do not need to switch models or change a setting, and the new versions replace the previous ones in the background. Muse is Meta's newer agent-style assistant, built to carry out multi-step tasks for a user, while Meta AI is the general assistant built into Meta's social and messaging apps. Both now run on models that inherit behaviour from Wonder without the full cost of the larger system.

Meta's numbers, and what they leave out

All figures in the post come from Meta's own evaluations. Meta says Muse is 2.4 times less likely to forget critical information. On its Active Memory evaluation, the new models used critical context correctly in 90% of tasks, compared with 76% for the previous models. For Active Focus, Meta reports a 67% improvement in completing everyday tasks where unexpected issues arise, but it did not publish the starting rate, so the size of the absolute change is unknown.

Meta also says the current Muse model scores 47% on its Personal Superintelligence Index, up from 38%. The post gives no methodology for that index or for the Active Memory test, and no third-party evaluation was published. Meta did not name the new models, give their sizes, say how they were trained beyond the use of Muse Wonder, or provide separate results for Meta AI. It also gave no schedule for how fast the update reaches every account and region, and it announced no new user controls for what the memory feature keeps.

Why train small models with a big one

The approach is a familiar one in the industry: a large, expensive model is used to train smaller ones that are cheap enough to serve to hundreds of millions of people. Meta AI runs inside WhatsApp, Instagram, Facebook and Messenger, so even modest gains in memory and task completion reach a very large audience, while the cost of running Wonder for every free request would be prohibitive.

Muse Wonder launched on 12 October 2026 through the Meta Model API at $4 per million input tokens and $20 per million output tokens with a context window of about 1 million tokens, according to launch coverage. Muse Wonder Full is available only in the top Muse Code plans. Bringing Wonder's skills into the free assistants is the step that turns a frontier model, sold mainly to developers, into something most of Meta's users actually touch.

Context and what to watch

The update lands in a heavy period of spending for Meta. In its third-quarter results the company kept its 2026 capital spending forecast at up to $145 billion, with Zuckerberg pointing to Muse and Muse Wonder as progress toward what he calls personal superintelligence. In October Meta also released Muse Spark Open as an open-weight model, and this month it bought Conductor, the app for running parallel coding agents.

Points to watch are whether Meta publishes the evaluations behind the index, whether outside testers see the same drop in forgetting, and how the stronger memory is handled under privacy rules in the EU, where memory features in assistants draw regulator attention. More model coverage is in AI models.

For competitors, the move raises the floor for free assistants. Memory across sessions has become a standard selling point for consumer chatbots, and persistence on multi-step tasks is the main difference between a chat assistant and an agent. If Meta's numbers hold up outside its own tests, the free tier of an assistant built into apps used by billions of people now carries features that rivals often reserve for paid plans. That is the claim outside evaluators will want to check first.

Questions people ask

Do free Meta AI users get Muse Wonder?

No. They get specialized models trained using Muse Wonder that carry its Active Memory and Active Focus features. Muse Wonder itself stays with subscribers and developers, according to Meta.

What is Active Memory in Muse?

Meta's name for the model's ability to keep and retrieve critical details across conversations. Meta says the new models used critical context correctly in 90% of tasks, up from 76%.

Are Meta's Muse benchmark numbers independently verified?

No. The 2.4x, 67% and Personal Superintelligence Index figures come from Meta's own evaluations, and Meta has not published their methodology.

Sources

1. Bringing Specialized Models to Muse and Meta AI (Meta, 6 Nov 2026)
2. Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents (Meta, 12 Oct 2026)
3. Meta Launches Muse Wonder and Muse Wonder Full: Benchmarks, Pricing and How to Use It (MarkTechPost, 13 Oct 2026)
4. Meta Launches Muse Wonder, Its First Frontier AI Model, at $4 per Million Tokens (Technology.org, 13 Oct 2026)