Meta releases Muse Wonder models for agents
Meta has launched Muse Wonder and Muse Wonder Full and calls them its most capable models yet, built for agents that work across users' messages, calendars and documents.
Meta announced the Muse Wonder series on Friday. The two models, Muse Wonder and Muse Wonder Full, build on the company's earlier Muse Spark model. Meta says the release follows a 14-month overhaul of its AI efforts and that the models deliver state-of-the-art performance on personal and professional agentic tasks. The announcement does not include benchmark figures or say how the two variants differ.
Meta is pitching the models at expert domains, including law, finance and health, and says they also offer frontier-level coding performance. It highlights two capabilities:
- Active Memory carries relevant details across conversations.
- Active Focus lets the model keep working on a hard task: it tries other approaches, recovers from errors, asks for clarification and checks in before taking important actions.
On safety, Meta says it trained the models not to share private information without permission and to refuse requests that would cause real harm. It says it kept their ability to help developers and security teams find flaws in their own code. Meta says the models were trained, evaluated and released under its Meta Superintelligence Scaling Framework. It says a full report with additional safety evaluations will be published "soon."
Muse Wonder is available now through the Meta Model API and Muse Code. Meta says it will soon power more of Muse, starting with subscribers. The announcement gives no pricing, context-window or availability details for Muse Wonder Full. Meta points readers to its research blog for more on how the models were built and tested.