At a glance
- Muse Wonder and Muse Wonder Full were introduced October 12.
- Meta emphasizes persistent context, sustained task execution and coding.
- Access begins through developer tools, with the Muse rollout to follow.
Meta introduced Muse Wonder and Muse Wonder Full on October 12, describing the pair as its most capable artificial intelligence models so far. The launch is aimed at personal and professional agents that handle tasks over time, rather than only answer individual questions. Reuters reported that the release intensifies Meta’s competition with OpenAI and Anthropic. [1] [3]
Memory and sustained work
Meta’s announcement centers on two capabilities. Active Memory is intended to retain important information across conversations and recover it when relevant. Active Focus is intended to keep a model working through obstacles, trying alternatives and asking for clarification when needed. The company says these features support tasks such as examining a lease, navigating an insurance claim and building software. [1]
Its research account explains how it tests sustained effort, naming FocusBench and SWE-Marathon. The latter evaluates work on software projects that require prolonged execution. Meta also describes evaluations in law, finance and health, alongside coding. These are the developer’s descriptions of its testing and intended uses, not independent proof that the models reliably complete every task in those fields. [2]
Where the models are available
Muse Wonder was made available through the Meta Model API and Muse Code at launch. Muse Wonder Full was offered to Muse Code subscribers. Meta said the models would subsequently power more experiences in Muse, beginning with subscribers. The distinction matters: a developer release does not mean the upgraded system is already available to everyone using the consumer assistant. [1]
Competition and unanswered questions
Reuters reported that Meta shares were up about 4% in afternoon trading after the announcement. That market response shows investor attention, but does not validate the technology. Meta’s research post presents benchmark performance and says a further safety report is planned. Independent experience with reliability, privacy and real-world task completion remains separate from the company’s launch claims. [2] [3]