Summary
Meta has introduced Muse Wonder and Muse Wonder Full, models it says are designed for personal and professional AI-agent tasks. Muse Wonder is available through the Meta Model API and Muse Code; Muse Wonder Full is available to Muse Code subscribers.
Meta introduced Muse Wonder and Muse Wonder Full on October 12, 2026, describing them as its most capable models for personal and professional AI-agent tasks. The company says Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers. Muse Wonder is also due to reach Meta’s Muse product, initially for subscribers.
The announcement presents the models as a step towards AI agents that can retain useful context, work through longer tasks and act across tools. Meta says both models were trained at its Prometheus data centre and build on work in pretraining and reinforcement learning.
Memory and sustained work
Meta calls one part of the system Active Memory: carrying relevant information between conversations and retrieving it when it helps with a later task. Active Focus refers to continuing towards a goal by trying different approaches, responding to new information and working through obstacles. Meta says the models can ask for clarification and check in before important actions.
These capabilities address two practical challenges for agents. A system working across a series of steps may need to remember earlier instructions, and it may need to adapt when an initial approach fails. Meta’s examples include planning a run around personal preferences, making a wedding seating chart from complex requirements, and building a playable 3D kitchen game from image inputs. These are company demonstrations of intended uses, rather than a description of independent testing.
Meta says its evaluations cover expertise in areas including law, finance and health, alongside memory, sustained task execution and coding. It cites Harvey’s Legal Agent Benchmark, which tests models on source documents and detailed instructions, FocusBench for focus, and SWE-Marathon for work on long-running software projects. The company also says it trained the models on diverse, long-running coding tasks involving repeated changes, tests and revisions.
Generalising beyond targeted tasks
The announcement describes an analysis comparing Muse Wonder Full with Muse Spark 1.3 as reasoning effort at inference time increases. Meta reports that Wonder Full scored better on unfamiliar tasks than the trend from Spark 1.3 would predict at a comparable level on tasks targeted during post-training. The company attributes this pattern to stronger base models, arguing that they allow post-training gains to carry over more effectively to tasks the models were not specifically trained to target.
That distinction matters for personal agents, which may encounter combinations of goals, tools and constraints that are difficult to anticipate in advance. Meta characterises the Muse Wonder series as state of the art across key benchmarks for agents, memory and coding. The announcement describes evaluation areas but does not provide individual benchmark scores in the supplied text.
Availability and safety claims
Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full is currently available for Muse Code subscribers. Meta says Muse Wonder will soon power experiences in Muse and support artifact generation and processes requiring more intelligence; the company says the Muse rollout will begin with subscribers.
Meta also says it trained the models to resist adversarial instructions and handle private information carefully. It describes Muse Wonder as its most aligned model to date and says the family has its strongest safeguards so far against chemical, biological, radiological and nuclear (CBRN) and cyber misuse, while retaining access to frontier cyber capabilities for defensive use. The company says a fuller report with additional safety evaluations will be available soon.