Meta launches Muse Wonder models for long-running agent work

Meta launched Muse Wonder and Muse Wonder Full on October 12. Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is limited to Muse Code subscribers. That gives developers a new preview model for agents that work across longer tasks, though the Full model is not available through the API.

Meta’s developer documentation lists the standard model as muse-wonder-1.0 with a 1,048,576-token context window. The Muse Wonder product page says Muse Wonder Full comes with two paid Muse Code subscription tiers.

In its launch announcement, Meta described two features intended to support work over time. Active Memory is designed to carry relevant user context between conversations and retrieve details for the task at hand. Active Focus is meant to keep the models pursuing a goal, try different approaches, recover from errors, ask for clarification and check in before important actions. These are Meta’s descriptions; the opened evidence does not independently show how the features perform in real-world deployments.

Meta is positioning the family for personal and professional agentic tasks across software development and fields including law, finance and health. The company also calls the models state of the art and frontier-level, but its model card places important limits on those comparisons. Some results are internal, while others come from third parties or model providers. Externally reported coding scores can also use different agent harnesses. Meta says its own competitor runs are best-effort and may not be optimally tuned for proprietary models.

The model card does not disclose the models’ architecture, parameter count, training compute or training-data composition. It warns that the models can produce inaccurate, biased or objectionable responses, make mistakes in multi-step reasoning, and perform differently depending on prompts, tools, serving setup and agent harness. Meta recommends use-case-specific testing and human confirmation before an agent takes irreversible actions.

On safety, Meta says it trained the models to avoid unauthorized disclosure of private information and harmful assistance while preserving defensive cybersecurity work. The company assessed Muse Wonder Full’s mitigated residual risk as “moderate or lower” across cybersecurity, chemical and biological, and loss-of-control domains. That is Meta’s assessment, not an independent safety certification. The company said a fuller report with additional safety evaluations will come later.