Meta launches Muse Wonder agent models via Meta Model API

Meta says the release builds on a ground-up overhaul of its AI efforts that began fourteen months ago and on Muse Spark, which outlined its approach to predictable, rigorous scaling. The Muse Wonder series was trained at Meta's gigawatt-scale Prometheus data center, and Meta describes the models as the next step on its scaling ladder toward personal superintelligence.

Beyond raw benchmark scores, Meta argues that a stronger base model makes post-training gains carry over more effectively to unfamiliar tasks. It compares Muse Wonder Full with Muse Spark 1.3, noting that given its in-distribution performance, Muse Wonder Full performs even better on out-of-distribution tasks than the Muse Spark 1.3 trend would predict. Meta frames that versatility as especially important for personal agents, which must handle combinations of goals, tools and constraints that cannot all be anticipated during training.

Meta also ties the release to safety, saying a personal agent may handle messages, calendars and private documents, which it says demands discretion with private information and resistance to misuse. It says Muse Wonder offers its strongest safeguards so far against CBRN and cyber misuse. Muse Wonder is also set to power experiences in Muse, first rolling out to subscribers, and to support artifact generation and specialized models later.