Meta has announced the Muse Wonder series, consisting of Muse Wonder and Muse Wonder Full, which the company describes as frontier models designed for personal and professional agent tasks. According to Meta’s announcement, these models represent a significant step toward what the company calls “personal superintelligence.”
The models are now available through the Meta Model API and Muse Code, with Muse Wonder Full accessible to Muse Code subscribers. Muse Wonder will soon power experiences in Muse and will roll out first to subscribers. The company plans to use Muse Wonder to develop specialized models and to power artifact generation and processes requiring increased intelligence.
Muse Wonder brings together two core capabilities: Active Memory, which tracks important context across conversations and retrieves relevant details for specific tasks, and Active Focus, which enables the model to sustain effort toward challenging goals by trying alternative approaches, working through obstacles, and adapting to new information.
According to Meta, the models achieve state-of-the-art results across key benchmarks for personal and professional agents, memory, and coding. The company evaluated Muse Wonder and Muse Wonder Full across several specialized domains. In legal work, the models were tested on Harvey’s Legal Agent Benchmark, which requires models to work with source documents and follow detailed instructions to produce legal work products. The models were also evaluated on tasks in finance and health that test their ability to apply specialized knowledge to problems people encounter at work and in daily life.
For coding capabilities, Meta notes that Muse Wonder and Muse Wonder Full deliver frontier coding performance, supporting projects from small applications to complex software. The evaluation included FocusBench and SWE-Marathon, which tests work on long-running software projects.
Meta trained both models at its Prometheus data center using a gigawatt-scale infrastructure. The company built the models on larger, stronger base models than its previous Muse Spark 1.3, which the company says yields post-training gains that transfer more effectively to unfamiliar tasks.
Regarding safety, Meta states that Muse Wonder is its most aligned model to date and offers strong safeguards against CBRN (chemical, biological, radiological, and nuclear) and cyber misuse while preserving access to frontier cyber capabilities for defensive use. The company emphasizes that alignment is central to how it builds personal superintelligence, particularly given that personal agents take actions on a user’s behalf and may handle private information such as messages, calendars, and documents.
Key facts
- Muse Wonder and Muse Wonder Full are available through the Meta Model API and Muse Code
- The models combine Active Memory to track context and Active Focus to sustain effort toward goals
- Both models achieve state-of-the-art results on benchmarks for agents, memory, and coding
- The models were trained at Meta’s gigawatt-scale Prometheus data center
- Muse Wonder is described as Meta’s most aligned model with strong safeguards against CBRN and cyber misuse