Meta is rolling out new models trained using its frontier Muse Wonder model to all users of its MUSE personal-agent app and Meta AI, starting November 81,3. Subscribers retain access to Wonder itself for harder tasks.
The company describes the new models as specialized versions of its existing ones, trained using Muse Wonder on behaviors that matter in a consumer app rather than built as smaller general-purpose variants of the frontier model. Meta has not named the new models and has not detailed its method for using a large model to train smaller ones.
Task-specific tuning over general capability
Meta says it tuned the models for specific use cases such as booking a restaurant table, managing an inbox, and recovering when a website breaks mid-task. The company trained them on a variety of real-world bottlenecks so the agents can retry, recover, or find alternatives on a user's behalf. The agents can also take corrections in the middle of a task without losing track and decide when to work autonomously versus when to check with the user.
Meta reports a 67% improvement in finishing routine tasks when the new models encounter unexpected problems inside MUSE, though the company did not share the baseline for that figure. An internal evaluation found the new models correctly use critical context in 90% of tasks, up from 76% for the previous model.
Active Memory and Active Focus reach free users
The rollout brings two features first introduced with Wonder to all MUSE and Meta AI users: Active Memory and Active Focus. Active Memory helps models retain details from earlier conversations and carry them into new ones. Active Focus keeps models working on a task when something goes wrong.
Meta originally launched Wonder and the more capable Wonder Full on October 12. Wonder is available through the Meta Model API and in Muse Code, while Wonder Full is limited to Muse Code's two most expensive plans. On the API, Wonder costs $4 per million input tokens and $20 per million output tokens.
MUSE launched in September on the smaller Muse Spark model and has a free tier, giving it a lower barrier to entry than OpenAI's Dots or Anthropic's DUET, which are available only to subscribers on those vendors' highest pricing tiers.
Separately, Meta announced on November 6 that Spaces are now available everywhere MUSE is available, letting users create persistent, customized experiences inside the app by describing an idea and refining the result with MUSE2. Spaces can support routines, goals, interests, practical work, or entertainment, and users can schedule when particular spaces appear in the interface.
ANALYSIS The distillation approach lets Meta extend frontier-grade behaviors to its entire free user base without serving the full Wonder model at scale, keeping the heavier model as a subscriber differentiator.