Meta is rolling out specialized versions of its existing AI models, trained with its Muse Wonder series, to Muse and Meta AI users, the company said in a blog post1.
The release extends two features, Active Memory and Active Focus, from Muse Wonder to the models behind Meta's consumer products. The models are not Muse Wonder itself. Meta describes them as specialized versions of existing models that were trained using it.
Meta's Muse Wonder launch post2 said Muse Wonder was available through the Meta Model API and Muse Code, and that Muse Wonder Full was available to Muse Code subscribers. It said Muse Wonder would soon power experiences in Muse, first for subscribers. It also said Meta would later use Muse Wonder to develop specialized models for everyone using Muse. The new post follows that plan. It does not say whether Muse Wonder itself now powers Muse.
Meta said Active Memory trains the models to recognize which information is temporary and which is permanent, and to separate what is still relevant from what has become stale. As an example, it said its improved models in Muse "show a 2.4x reduction in forgetting critical information necessary for helping you achieve your goals."
Active Focus, which Meta describes for Muse, is meant to keep a model working toward a goal despite unexpected obstacles. Meta said it trained the models on a variety of real-world bottlenecks. It said its models in Muse show a 67% improvement in completing everyday tasks where unexpected issues arise. According to the post, the Muse models also adapt better to steers and clarifications in the middle of running tasks.
For Meta AI, the post gives no figures. It says the same improvements make the assistant a better thinking partner. Meta said it can break complex problems into steps, ask sharper follow-up questions and keep up when a user changes direction mid-conversation.
Both figures come from Meta's post alone. It names no benchmark, baseline or test method, and both numbers are stated for Muse. The Muse Wonder post's description of that model as state-of-the-art rests on Meta's own evaluations, which compare it with other frontier models.
The specialized-models post carries no date. A Techmeme post summarizing a New Stack report on the announcement is timestamped Nov. 7 UTC4. The post includes no independent test results, pricing, regions, rollout schedule, architecture changes or safety evaluations. The Muse Wonder post said a full report with additional safety evaluations would be available soon.