Meta rolls out Muse Wonder-based models with memory and focus upgrades

Meta has launched specialized versions of its existing models trained on Muse Wonder, rolling them out to all Muse and Meta AI users. The series was released last month, focusing on fast, cutting-edge performance, and includes Muse Wonder and the full Muse Wonder Full.

The update brings two key capabilities: Active Memory lets the model keep key user info across chats and recall it when relevant, cutting the problem of forgetting goal-related info in Muse by 2.4x. Active Focus keeps the model pushing toward its goal even when surprises happen, lifting task completion by 67%.

Meta's internal tests show Muse now uses key context correctly in 90% of tasks, up from 76%, and its Personal Superintelligence Index score rose from 38% to 47%. The numbers are self-reported, with no independent third-party review yet.

Developers can access the models via Meta Model API and Muse Code, while Muse's new models are first open to subscribers paying $20 or $100 a month. The series is closed-source.