On Nov. 6, Meta said it is using Muse Wonder — the frontier agent models it introduced in mid-October — to train specialized versions of the models that already power Muse and Meta AI. Those specialized models are rolling out to people who use both products, and they now include Active Memory and Active Focus, two capabilities that first appeared with Wonder. Meta AI Research, "Bringing Specialized Models to Muse and Meta AI" (Nov. 6, 2026) Meta Newsroom, "Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents" (Oct. 12, 2026)
That is a quieter announcement than a new flagship model name, and a more important one for most users. Wonder itself is still a developer and subscriber system. Muse Wonder is available through the Meta Model API and Muse Code, and Muse Wonder Full is available to Muse Code subscribers. The October launch said Wonder would later power more of the consumer Muse app, starting with subscribers. The November post is the broader follow-through: Meta is using Wonder to train the models Muse and Meta AI already run, rather than putting Wonder itself in front of everyone. Meta AI Research, "Bringing Specialized Models to Muse and Meta AI" (Nov. 6, 2026) Meta Newsroom, "Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents" (Oct. 12, 2026) Frederic Lardinois, The New Stack, "Meta's best AI model now trains the one most of its users get" (Nov. 6, 2026)
Frederic Lardinois of The New Stack, writing the same day, reported that starting Friday, everyone using Meta's Muse agent and Meta AI gets the new Wonder-trained models, while subscribers still get Wonder itself for harder tasks. Meta has not named the specialized models. Frederic Lardinois, The New Stack, "Meta's best AI model now trains the one most of its users get" (Nov. 6, 2026)
What Meta says changed
Active Memory is meant to carry useful details from one conversation to the next and notice when something said earlier becomes relevant again. Meta said it trained the models to tell temporary information from permanent information, to reason across time, and to separate what is still relevant from what has gone stale. On the company's own Active Memory evaluation, the current Muse model correctly used critical context in 90% of tasks, compared with 76% for the previous Muse model. Meta also reported a 2.4-times reduction in forgetting information it considers critical to helping someone reach a goal. Those figures come from Meta's internal tests, not from an independent benchmark. Meta AI Research, "Bringing Specialized Models to Muse and Meta AI" (Nov. 6, 2026) Frederic Lardinois, The New Stack, "Meta's best AI model now trains the one most of its users get" (Nov. 6, 2026)
Active Focus is the counterpart for when a task goes sideways. Meta said it trained the models on "a variety of bottlenecks reflective of the real world" so they can try harder, recover, or find another route. In Muse, the company reported a 67% improvement in completing everyday tasks when unexpected issues come up. Lardinois noted that Meta did not publish the baseline completion rate behind that gain. Meta AI Research, "Bringing Specialized Models to Muse and Meta AI" (Nov. 6, 2026) Frederic Lardinois, The New Stack, "Meta's best AI model now trains the one most of its users get" (Nov. 6, 2026)
Meta also published an internal Personal Superintelligence Index score of 47% for the current Muse model, up from 38% for the previous one. The announcement shows those two scores and does not spell out how the index is built. Meta AI Research, "Bringing Specialized Models to Muse and Meta AI" (Nov. 6, 2026)
The specialized models are not described as a smaller, general-purpose copy of Wonder. Meta said they were trained to pick up behaviors that matter in a consumer agent: knowing when to keep working independently and when to collaborate live with the user, and adapting when someone steers or clarifies in the middle of a task without losing the thread. In Meta AI, the company said the same improvements should make the assistant better at breaking problems into clear steps, asking sharper follow-up questions, and keeping up when a person changes direction mid-conversation. Meta AI Research, "Bringing Specialized Models to Muse and Meta AI" (Nov. 6, 2026) Frederic Lardinois, The New Stack, "Meta's best AI model now trains the one most of its users get" (Nov. 6, 2026)
Two products, two price points
The split Meta is drawing is familiar even if the training method is only partly explained. On the API, The New Stack reported that Wonder costs $4 per million input tokens and $20 per million output tokens. Muse has a free tier, which Lardinois argued gives it a lower barrier than OpenAI's Dots or Anthropic's Duet, products he said are limited to those companies' highest subscription tiers. Meta AI, meanwhile, runs inside WhatsApp, Instagram, Facebook and Messenger. Training lighter-weight models with Wonder is likely cheaper than serving Wonder itself to that many people, he wrote. Meta has not detailed how it used Wonder to train the specialized models. Meta Newsroom, "Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents" (Oct. 12, 2026) Frederic Lardinois, The New Stack, "Meta's best AI model now trains the one most of its users get" (Nov. 6, 2026)
When Wonder launched on Oct. 12, Meta described it as built for real-world agent work such as navigating an insurance claim, understanding a lease, or making sense of someone's finances, along with coding. It also said the Wonder family was trained to avoid sharing private information without permission, refuse help with actually harmful acts, and still allow developers and security teams to use the models to find flaws in their own code. The company said it trains and releases its most capable models in line with its Meta Superintelligence Scaling Framework, and that a fuller safety report would be available later. Meta Newsroom, "Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents" (Oct. 12, 2026)
None of the November consumer metrics have been independently replicated in the materials Meta published, and the research post does not say how quickly the new models will reach every account or every region. The New Stack's report that the rollout started Friday is the most specific public timing so far. The cautious reading is that Meta is shipping Wonder-trained memory and recovery behavior into the products most people already open, and that users will have to judge those claims in daily use. Meta AI Research, "Bringing Specialized Models to Muse and Meta AI" (Nov. 6, 2026) Frederic Lardinois, The New Stack, "Meta's best AI model now trains the one most of its users get" (Nov. 6, 2026)
The Louisiana story is the cluster, not this model card
There is a Louisiana story here, but it is not that these specialized Muse models were trained in Richland Parish. Meta has not said where the new consumer models were trained. What the company has said, separately, is that it is building the industrial side of the same AI push in northeast Louisiana.
On July 13, Meta said it is expanding its AI-optimized data center in Richland Parish to 5 gigawatts of compute capacity, the largest site in its fleet, and that the campus will house Hyperion, its largest multi-gigawatt AI training cluster. The company called the expansion an investment of more than $50 billion in the region. Meta Data Centers, "Deepening our investment in Richland Parish, Louisiana" (July 13, 2026)
Meta said the project is expected to support over 7,500 jobs at peak construction and 1,000 operational roles once the site is running. It also said it has already contracted more than $1.6 billion with local Louisiana businesses. In the same post, Meta said an energy agreement with Entergy is expected to save Entergy Louisiana customers $2.65 billion over 20 years, a total it broke down as about $2 billion from the recent agreement on top of $650 million announced earlier. The recent agreement, Meta said, will fund seven new natural-gas generating plants, three grid-scale batteries, nuclear uprates and other purchased power. Meta Data Centers, "Deepening our investment in Richland Parish, Louisiana" (July 13, 2026)
That is the scale of infrastructure sitting behind consumer features such as remembering what someone said last week or recovering when a booking page breaks. The models reaching Muse and Meta AI and the cluster rising in Richland Parish are not the same project, and no public document ties this week's specialized models to Hyperion. They are two public faces of one company's bet that personal agents will need both better everyday software and far more training capacity than older data centers were built to provide.
Sources
- Meta AI Research, "Bringing Specialized Models to Muse and Meta AI" (Nov. 6, 2026) Meta AI Research, "Bringing Specialized Models to Muse and Meta AI" (Nov. 6, 2026) — Meta AI Research, "Bringing Specialized Models to Muse and Meta AI" (Nov. 6, 2026)
- Meta Newsroom, "Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents" (Oct. 12, 2026) Meta Newsroom, "Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents" (Oct. 12, 2026) — Meta Newsroom, "Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents" (Oct. 12, 2026)
- Frederic Lardinois, The New Stack, "Meta's best AI model now trains the one most of its users get" (Nov. 6, 2026) Frederic Lardinois, The New Stack, "Meta's best AI model now trains the one most of its users get" (Nov. 6, 2026) — Frederic Lardinois, The New Stack, "Meta's best AI model now trains the one most of its users get" (Nov. 6, 2026)
- Meta Data Centers, "Deepening our investment in Richland Parish, Louisiana" (July 13, 2026) Meta Data Centers, "Deepening our investment in Richland Parish, Louisiana" (July 13, 2026) — Meta Data Centers, "Deepening our investment in Richland Parish, Louisiana" (July 13, 2026)