Summary
Meta has introduced Muse Wonder and Muse Wonder Full, models designed for personal and professional agent tasks. The company describes new memory and sustained-task capabilities and has begun offering access through Meta Model API and Muse Code.
Meta introduced Muse Wonder and Muse Wonder Full on October 12, 2026, describing them as its most capable models to date for personal and professional agent tasks. The company says Muse Wonder combines systems it calls Active Memory and Active Focus to retain relevant context and keep working toward a user’s goal. Muse Wonder is available through Meta Model API and Muse Code; Muse Wonder Full is available to Muse Code subscribers.
Memory, sustained work and coding
Meta uses Active Memory to describe carrying useful context across conversations and retrieving relevant details for a task. Active Focus refers to continuing to work through a difficult task, trying different approaches, adapting to new information and asking for clarification or checking in before important actions.
The company reports results across evaluations for law, finance and health, as well as memory, focus and coding. One legal evaluation it names, Harvey’s Legal Agent Benchmark, tests whether a model can use source documents and follow detailed instructions to produce legal work. Meta also cites FocusBench and SWE-Marathon, which it uses to assess focus and sustained work on long-running software projects. It says the models were trained on varied, long-running coding tasks involving requirements, code changes, testing and revision.
Meta also describes an analysis of performance on tasks targeted during post-training and on unfamiliar tasks. In comparisons at different levels of test-time computation—the processing used while a model works through a task—the company says Muse Wonder Full performs better on unfamiliar tasks than the trend from Muse Spark 1.3 would predict for its performance on targeted tasks. Meta attributes this to the newer models’ larger, stronger base models, which it says help post-training improvements carry over to unfamiliar problems. This matters for agents that may encounter combinations of goals and constraints not anticipated during training.
Rollout and safeguards
Meta says Muse Wonder will soon power features in Muse, with subscribers first in line. The company also plans to use it for artifact generation and processes requiring greater intelligence, and says it intends to develop specialised models based on Muse Wonder for broader use in Muse.
Because an agent may handle messages, calendars and private documents, Meta says it trained the models to resist adversarial instructions and handle private information carefully. It describes Muse Wonder as having its strongest safeguards so far against chemical, biological, radiological and nuclear misuse and cyber misuse, while retaining frontier cyber capabilities for defensive use under its Meta Superintelligence Scaling Framework. Meta says a fuller report with additional safety evaluations will be available soon.
The performance and safety descriptions are Meta’s account of its own evaluations. The announcement names evaluation areas and benchmarks but does not include numerical results in the supplied text.