Meta launches Muse Wonder models

Meta said it had spent 14 months rebuilding its AI work and had now launched the Muse Wonder series, which included Muse Wonder and Muse Wonder Full. It said the new models were built on Muse Spark and were its strongest systems so far.

The company said the models were aimed at agentic tasks in personal and professional settings. It pointed to work such as handling insurance claims, reading leases, understanding finances and coding for games or new applications, and said they were also intended for law, finance and health.

Meta said the models included Active Memory, which carried relevant details across conversations, and Active Focus, which let them keep working when tasks became difficult, try other approaches and recover from errors. It said they would ask for clarification and check before carrying out important actions.

On safety, Meta said it trained the family to avoid disclosing private information without permission, to refuse requests for harmful acts and to support defensive security work. It said the models were trained, evaluated and released under its Superintelligence Scaling Framework, and that a fuller safety report would follow soon.

Meta said Muse Wonder was available through the Meta Model API and Muse Code. It said the model would soon power more of Muse, with the rollout beginning for subscribers, and described the release as a step towards personal superintelligence.