Meta (aka Facebook) has released a pair of new AI models called Muse Wonder and Muse Wonder Full, which the company describes as its most capable systems yet for personal and professional tasks. The announcement marks Meta's first entry into the frontier model category and represents fourteen months of rebuilding the company's AI stack from first principles.
Key capabilities and benchmarks
The Muse Wonder series introduces two technical features that Meta calls Active Memory and Active Focus. Active Memory allows the model to carry important context about a user across multiple conversations and retrieve relevant details when needed. Active Focus describes the model's ability to continue working toward difficult goals over time - trying different approaches, working through obstacles, adapting to new information, and asking for clarification when appropriate.
According to Meta's evaluation data, the models achieve state-of-the-art results on several benchmarks.
The models were trained at Meta's gigawatt-scale Prometheus data center and build on larger base models than previous Muse systems. Meta notes that this stronger foundation helps post-training improvements transfer more effectively to unfamiliar tasks, which matters for personal agents that encounter unexpected combinations of goals and constraints. Muse Wonder and Muse Wonder Full demonstrate this through better out-of-distribution performance relative to their in-distribution scores.
Muse Wonder is available immediately through the Meta Model API and Muse Code, while Muse Wonder Full is restricted to Muse Code subscribers. Meta plans to expand access to Muse subscribers and use the models to develop specialized systems for broader distribution.
On safety, Meta reports that Muse Wonder offers the company's strongest safeguards to date against misuse for chemical, biological, radiological, nuclear, and cyber threats, while preserving defensive cyber capabilities. The model is also designed to handle private information carefully and resist adversarial instructions. A full safety report is expected soon.
This release is significant because it represents Meta's first frontier-class model and signals a major commitment to AI agents that can work autonomously on behalf of users over extended periods. The focus on memory and sustained attention addresses two of the main limitations of current AI assistants, which typically forget context between sessions and struggle with long-running tasks. If the benchmark results translate to real-world performance, Muse Wonder could establish Meta as a serious competitor in the enterprise and personal agent markets.