Meta ships Muse Wonder and keeps Muse Wonder Full inside Muse Code

On October 12, 2026, Meta introduced Muse Wonder and Muse Wonder Full as its most capable models to date for personal and professional agent tasks. For builders, the split is who can call which variant today. Muse Wonder is available through the Meta Model API and in Muse Code. Muse Wonder Full is available for Muse Code subscribers. Unlike Wonder, the posts do not list Full on the Model API.

Meta does not say the Muse product runs on Wonder today. It says Wonder will soon power experiences in Muse, rolling out first to subscribers. The research post adds that Wonder will also power artifact generation and processes that require more intelligence, and that Meta plans to use it to develop specialized models for people using Muse.

Meta's examples include navigating an insurance claim, making sense of a lease, and making sense of finances or health information, plus coding from a game or a small business app to more complex software. It places that work in law, finance, health, and coding, on groundwork from Muse Spark, after an AI overhaul it dates to fourteen months before this release. Training, it says, ran at its gigawatt-scale Prometheus data center.

Active Memory is Meta's name for carrying what matters across conversations and bringing back details relevant to the task. Active Focus is its name for staying on a hard goal: trying other approaches, recovering from errors, working through blockers, adapting to new information, asking for clarification, and checking in before important actions. Coding, the research post says, brings those capabilities together: tracking requirements, changing a codebase, and repeating cycles of testing and revision. Both models were trained on diverse, long-running tasks, with FocusBench and SWE-Marathon named as evaluations of focus and sustained software work.

Meta claims state-of-the-art results on agent, memory, and coding benchmarks, and the research post shows charts for suites such as Harvey's Legal Agent Benchmark, Vals Finance Agent V2, and MemoryBench. The scores are in images, not in the prose, so these posts alone do not settle a ranking.

Meta says the family was trained to avoid sharing private information without permission, to turn down requests to assist in performing actually harmful acts, and to strengthen defenses against cyberattacks, while still letting developers and security teams find flaws in their own code. The research post calls Muse Wonder its most aligned model to date and says it offers the strongest safeguards so far against chemical, biological, radiological, nuclear, and cyber misuse, with defensive cyber use preserved under the Meta Superintelligence Scaling Framework. A full report with additional safety evaluations is still forthcoming.

Read the newsroom post for who can call which variant, then the research blog for the definitions of Active Memory and Active Focus.

Sources