Meta’s best AI model now trains the one most of its users get

Meta has introduced new AI models for its Muse agent and Meta AI, which are now training the version most users interact with. These models, dubbed "specialized" versions of the existing ones, were trained using Meta's frontier model, Muse Wonder. Subscribers still have access to Wonder for more demanding tasks.

The new models come with two features initially introduced with Wonder: Active Memory and Active Focus. These features enable the models to remember details from earlier conversations and adapt when faced with interruptions or errors during a task. For instance, Active Memory helps the models distinguish between temporary and permanent information, while Active Focus enables them to recover from bottlenecks and find alternative solutions.

Meta launched Wonder on October 12, offering it through the Meta Model API and in Muse Code, its coding agent. Wonder Full, a more capable version, is limited to the two most expensive plans of Muse Code. Wonder was made available to everyone in Muse, Meta's personal agent app, in September. Initially running on the smaller Muse Spark model, Muse has gained popularity as a breakthrough for Meta, providing consumers with a personal agent at a lower cost than competitors like OpenAI's Dots or Anthropic's Duet.

The key difference with these new models is that Meta trained them on specific behaviors relevant to a consumer app like Muse. This allows agents to determine when to work independently and when to seek user input, as well as recover from errors mid-task without losing context. Meta hasn't disclosed the specifics of its training method, but the models are specifically tuned for tasks such as booking a restaurant table and managing a user's inbox.

Active Memory and Active Focus features improve the models' ability to retain context and maintain task completion rates, with a 90% accuracy rate for correctly using critical context in tasks compared to 76% for the previous model.