The new models are “specialized” versions of Meta’s existing models, the company said, and were “trained using Muse Wonder,” its frontier model. Starting Friday, they are available to everyone using the Muse agent app and Meta AI across WhatsApp, Instagram, Facebook and Messenger.
Muse, which launched in September on the smaller Muse Spark model, became somewhat of a breakout hit for Meta and introduced consumers to the idea of personal agents. Unlike many personal agents, Muse has a free tier, lowering the barrier to entry compared with services such as OpenAI’s Dots or Anthropic’s Duet, which are restricted to subscribers on those vendors’ highest pricing tiers.
Active Memory helps the models remember details and take them from earlier conversations into new ones. Meta wrote that it had trained its models to recognise which information is temporary and which is permanent, to reason across time, and to distinguish between what is still relevant and what has become stale. On the company’s internal evaluation, the new models in Muse correctly used critical context in 90 percent of tasks, up from 76 percent for the previous model.
Active Focus is designed to keep models working on a task when something goes wrong. Meta said it trained the models “on a variety of bottlenecks reflective of the real world, so that they can try harder, recover, or find alternatives on your behalf.” In Muse, the company reported a 67 percent improvement in finishing routine tasks when these models encountered unexpected problems, though it did not share the baseline. The models are also better at deciding when to work autonomously and when to check with the user, and they handle mid-task corrections.
Meta has not detailed the training method, but said it trained the models on behaviours that matter in a consumer app rather than simply developing a smaller, general-purpose version of Wonder. Using a large model to train smaller ones is a common strategy for bringing frontier-level behaviour to consumer-scale products.