Meta released Muse Wonder and Muse Wonder Full on October 12, putting its strongest models directly into agent work. The important part is not another leaderboard shuffle. Meta is building a closed model around memory, long tasks and the private context a personal agent needs—and moving away from the open-weight identity that made Llama matter.
Wonder is built for jobs that refuse to fit in one prompt
In its launch report, Meta describes two ideas that sound simple and are hard to make reliable. “Active Memory” retrieves details from earlier conversations when they become relevant. “Active Focus” keeps a model working through blockers, tries another route and asks before consequential actions. Those are agent features, not prettier chat.
The demonstrations are deliberately ordinary: plan a run around a person’s circumstances, arrange a wedding seating chart, build a small game and prepare custom merchandise. That is sensible positioning. A model that can remember your constraints and keep a multi-step job moving is more useful than one that wins a riddle and forgets why you opened it.
Meta says the models lead on several agent, memory and coding evaluations. Those are Meta’s results, and its full safety report is still “coming soon.” Treat the launch tables as a strong reason to test Wonder, not as a verdict delivered by the company being tested.
The distribution is split
Muse Wonder is available through the Meta Model API and Muse Code. Wonder Full is restricted to Muse Code subscribers. The standard model is also due to reach paid Muse subscribers first. That makes the product story more complicated than “Meta caught up”: the most capable version is not a downloadable Llama successor, and access depends on which Meta surface you use.
This closed release changes the trade. Llama gave teams weights they could inspect, adapt and run on their own infrastructure. Wonder gives Meta a model it can improve alongside a consumer agent, its developer API and billions of existing accounts. Buyers get convenience and a potentially strong price; Meta keeps control of the model and the relationship.
The real test is permission, not memory
Personal context becomes valuable only when the agent can act on it. It also raises the cost of a mistake. Meta says Wonder asks for clarification and checks before important actions, and says it has trained the family to handle private information carefully. Those promises should be tested at the product boundary: what can the agent read, what can it change, what needs approval, and what remains in the activity log?
That is the same question raised by Muse’s earlier permission model. A more capable model does not make the approval screen less important. It makes every permission more valuable.
For developers, the useful first evaluation is one long task with changing requirements and a deliberate contradiction. Give Wonder a plan, change one fixed constraint, then see whether it repairs the plan or quietly breaks the rule. For everyone else, wait for the Muse rollout and inspect the controls before handing it messages, calendars or payment details. Frontier performance is a claim. Staying inside the job is the product.
Cover art: AI-generated editorial composition by TMRW.
