Meta launches Muse Wonder and Muse Wonder Full for multistep AI agents

New features target longer agent jobs

Meta says it aimed the models at longer, multistep jobs that people hand to an AI agent. It cited working through an insurance claim as one example, and said coding and specialist fields such as law, finance and health are strengths.

The models bring two new features. Active Memory carries context about a user from one conversation to the next and raises details when they become relevant. Active Focus is meant to keep the model on a task when it becomes difficult.

Training ran on Prometheus data center

Meta trained both models at its gigawatt-scale Prometheus data center. It built on pretraining and reinforcement learning methods first used for Muse Spark in April.

待解问题

• How well do the models perform on the insurance claim and other multistep tasks outside Meta's own examples?
• Are the claimed strengths in coding, law, finance and health supported by independent testing?

Key facts

• Meta released Muse Wonder and Muse Wonder Full, which it calls its most capable models yet for running agents on personal and professional tasks.