Meta released Muse Wonder and Muse Wonder Full today, introducing AI models designed to remember relevant details across conversations and keep working through complex tasks. Meta says the models support software development and tasks involving legal, financial and health information. Both were trained at Prometheus, the company’s data center operating at gigawatt scale.
Active Memory is Meta’s name for retaining important context about a user between conversations and retrieving it when a later task needs it. Active Focus is intended to help Muse Wonder pursue difficult goals by changing approaches, addressing obstacles and adjusting to new information. Meta says that process includes requesting clarification when necessary and checking with users before important actions.
For coding, Meta says both models can follow requirements, edit different parts of a codebase and continue through testing and revisions. The company trained them using varied tasks that required extended work, aiming to support everything from smaller apps to more complicated software projects. One demonstration shows Muse Wonder creating a playable kitchen game in 3D using images that users can customize.
Meta’s evaluations cover specialized work in law, finance and health, including Harvey’s Legal Agent Benchmark (LAB), which tests producing legal work from supplied documents under detailed instructions. It uses FocusBench to assess focus and SWE-Marathon to test continued execution on software projects over extended periods. Meta also compares Muse Wonder Full with Muse Spark 1.3 on tasks targeted during post-training and unfamiliar tasks, saying Muse Wonder Full exceeds the unfamiliar-task performance predicted by the older model’s trend.
Muse Spark 1.3 arrived in September with max reasoning aimed at difficult reasoning and agentic tasks, alongside improvements to usability in practice. Meta followed that earlier release with an October 2 account of mathematicians collaborating with Muse Spark on six research papers. The new models extend that emphasis on complex work into demonstrations such as preparing a wedding seating chart around complicated matching requirements and building a personalized running plan through browser and app interactions.
Meta says it trained the Muse Wonder family to resist adversarial instructions and handle private information carefully, citing agents’ potential access to messages, calendars and documents. It also claims stronger safeguards against chemical, biological, radiological, nuclear and cyber misuse, while retaining cyber capabilities for defensive purposes under its Meta Superintelligence Scaling Framework. The company says it will publish a fuller report with further safety evaluations soon.
