Summary
Meta has introduced Muse Wonder and Muse Wonder Full, models designed for personal and professional AI-agent tasks. The company highlights memory across conversations, sustained work on complex tasks and coding.
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 the models combine memory across conversations with the ability to stay engaged on difficult tasks, and reports strong results on benchmarks for agents, memory and coding.
Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers. Meta says Muse Wonder will roll out in its Muse app, first to subscribers, and will also support artifact generation and processes requiring more intelligence.
Memory and sustained work are central to the design
Meta calls one of the models’ features Active Memory: carrying important context about a person across conversations and retrieving details relevant to a later task. The aim is to let an agent use earlier information when it becomes useful, rather than treating each interaction as isolated.
The second feature, Active Focus, describes how the models work toward a goal over multiple steps. Meta says they can try alternative approaches, respond to new information, work through blockers, ask for clarification and check in before important actions. These capabilities matter for tasks that involve several constraints or stages, such as planning, document work or software development.
Meta’s evaluations include legal, finance and health tasks. It cites Harvey’s Legal Agent Benchmark, which tests whether models can use source documents and follow detailed instructions to produce legal work products. For sustained task work, Meta names FocusBench and SWE-Marathon, the latter testing work on long-running software projects. The company also says the models were trained on diverse, long-running coding tasks involving repeated cycles of changes and testing.
The announcement includes demonstrations of Muse Wonder working across a browser and apps to create a running plan using personal preferences, and generating a playable 3D kitchen game from customizable image inputs. These are examples of the kinds of workflows Meta is targeting, rather than a description of a single fixed product feature.
Meta’s account of benchmark performance and safety
Meta says the Muse Wonder series achieves state-of-the-art results across key benchmarks for personal and professional agents, memory and coding. It also describes a comparison between Muse Spark 1.3 and Muse Wonder Full at different levels of test-time compute—the additional computation used while a model reasons through a task.
In that analysis, Meta compares performance on tasks targeted during post-training with performance on tasks not targeted during that stage. The company says Muse Wonder Full performs better on unfamiliar tasks than the trend from Muse Spark 1.3 would predict at a comparable level of performance on targeted tasks. Meta attributes this to stronger base models, arguing that their gains carry over more effectively to unfamiliar problems. That breadth is relevant to agents, which may encounter combinations of goals, tools and constraints that were not anticipated during training.
Meta also says the models were trained to resist adversarial instructions and handle private information carefully. It describes Muse Wonder as having its strongest safeguards to date against chemical, biological, radiological and nuclear risks, as well as cyber misuse, while retaining capabilities for defensive cybersecurity work. A fuller report with additional safety evaluations is expected from the company.
The announcement presents these models as a step toward more capable personal agents: systems that can use context and pursue goals over time, rather than only respond to isolated prompts. Their practical value will depend on how reliably those abilities work across real tasks and how carefully agents handle sensitive information and consequential actions.