Meta introduced the Muse Wonder series, including Muse Wonder and Muse Wonder Full, describing the models as its most capable systems yet for personal and professional agentic tasks.
The new models combine what Meta calls Active Memory and Active Focus, capabilities designed to help AI keep track of relevant user context and continue working toward complex goals over extended periods.
Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers.
Meta also plans to bring Muse Wonder into Muse, initially for subscribers, and use the model to power artifact generation and other processes requiring higher levels of intelligence.
The company said Muse Wonder will also serve as a foundation for developing more specialized models that can eventually be deployed across Meta’s broader products.Meta trained
Muse Wonder and Muse Wonder Full at its gigawatt-scale Prometheus data center and built on its recent efforts to scale both pretraining and reinforcement learning.
Meta said the models deliver state-of-the-art results across evaluations covering personal agents, knowledge work, long-term memory, coding, business workflows, finance, law, and other professional tasks.
A major focus of the new architecture is Active Memory.
The system is designed to retain important information across conversations and retrieve relevant details when they become useful for a new task.
Meta views this capability as essential for personal and professional agents that need to understand user preferences, priorities, and previous interactions over time.
The second major capability, Active Focus, is designed to help the models maintain progress toward difficult objectives by trying alternative approaches, working through obstacles, and adapting as new information becomes available.
The models can also ask for clarification when needed and check with users before taking important actions.
Meta evaluated sustained execution using tests including SWE-Marathon, which measures performance on long-running software engineering projects.
The company said Muse Wonder and Muse Wonder Full also deliver frontier coding capabilities, allowing them to track requirements, make changes across codebases, and complete repeated testing and revision cycles over extended tasks.
Meta highlighted domain expertise as another major area of improvement.
The models are designed to apply specialized knowledge to problems people encounter in areas such as law, finance, and health, including tasks that require working from source documents and following detailed instructions.
Meta also emphasized the models’ ability to perform on unfamiliar tasks that were not specifically targeted during post-training.
According to the company, Muse Wonder Full showed stronger out-of-distribution performance than would have been predicted based solely on improvements seen in its earlier Muse Spark model.
Meta said that broader capability is particularly important for personal AI agents, which may encounter combinations of goals, tools, and constraints that cannot all be anticipated during training.
The company also demonstrated Muse Wonder operating directly in a browser and across applications, including an example in which the model created a running plan based on personal preferences and circumstances.
Safety and alignment are another major focus of the Muse Wonder series.
Meta said personal agents may have access to sensitive information such as messages, calendars, and private documents, requiring models to exercise discretion and resist adversarial instructions.
The company said Muse Wonder includes its strongest safeguards to date against chemical, biological, radiological, nuclear, and cyber misuse while maintaining access to defensive cyber capabilities.
Meta described Muse Wonder as another step toward its broader goal of developing personal superintelligence capable of assisting users across both everyday activities and professional work.
