Meta began rolling out new models to everyone who uses its Muse agent and Meta AI, which it built as specialized versions of models it already had and trained using Muse Wonder, the frontier model it launched in October. The new models give free users Wonder’s Active Memory, which carries details forward between conversations, and Active Focus, which keeps a task going when something breaks, while Wonder itself stays with paying subscribers for harder work. Meta says on its own evaluation the updated Muse uses critical context correctly in 90% of tasks, up from 76%, and is 67% better at finishing everyday tasks that hit unexpected problems, though it gave no baseline for that second figure. Meta didn’t give the new models names or explain how it used Wonder to train them.
