Meta launches Muse Wonder, its top AI model, for agent tasks

Meta is positioning its newest flagship models around agent tasks rather than general chat alone. Insurance claim procedures and lease document understanding are the examples given, and the models are described as bringing domain knowledge in law, finance and health. Safety training is also part of the package: Meta says the models are trained not to share personal information without permission, to refuse harmful requests, and to strengthen defenses against cyberattacks.

Active Memory and Active Focus are the work-oriented features. Active Memory retains information important to the user across a conversation and recalls details relevant to the current task. Active Focus keeps the model working toward a goal when a task gets difficult, trying other approaches and recovering from errors, while asking for confirmation before important actions.

For now, availability is split. Muse Wonder is offered to developers through Meta Model API and Muse Code, while Muse Wonder Full goes to Muse Code subscribers. Muse agent features powered by Muse Wonder are planned next and will start with subscribers.