Meta started rolling out specialized versions of its existing models for Muse and Meta AI on Friday. They were trained using Muse Wonder, the frontier model the company introduced on October 12. Paying subscribers still get Wonder itself for harder work.
The new models have no names. Meta calls them “specialized” versions of the existing ones and says nothing about the method. Using a large model to train smaller ones is standard practice across the industry. What is worth a look here is what they were trained on.
Two capabilities carried down from the frontier model
Muse Wonder shipped with two named features, and both are now in the small models.
Active Memory is meant to remember what matters and notice when something older becomes relevant again. Meta writes that it trained the models to tell temporary information from permanent, to reason across time, and to separate what is stale from what still counts. On Meta’s own eval, the new models use the critical context correctly in 90 percent of tasks, against 76 percent for the previous ones. The company frames that as 2.4x less forgetting.
Active Focus covers the case where something breaks. Meta trained on bottlenecks modeled on real-world failures so the model tries harder, recovers, or finds another route. In Muse it reports a 67 percent improvement in completing everyday tasks where unexpected problems come up. It does not give a baseline for that 67 percent.
On Meta’s in-house Personal Superintelligence Index, the score goes from 38 to 47 percent. Meta also claims better judgment about when to push on alone and when to check in, plus the ability to take a correction mid-task without losing the thread.
Who gets to see this
The difference from the competition is reach. Muse launched in September on the smaller Muse Spark model and has a free tier. OpenAI’s Dots and Anthropic’s Duet are limited to the top subscription tiers. Meta AI also runs inside WhatsApp, Instagram, Facebook and Messenger. Wonder itself costs $4 per million input tokens and $20 per million output tokens on the API — at the user numbers Meta is talking about here, a distilled model is the only affordable option.
Memory is becoming a product category
This is the second memory story in two days. Yesterday Anthropic published its own account of what agent memory is made of and how Claude consolidates while idle. Today Meta puts numbers on the same problem inside a consumer product.
What stands out is that neither company treats remembering as a context-window problem any more. It is a training objective. Meta trains explicitly on the distinction between temporary and permanent. For developers the situation stays unsatisfying, because Meta neither discloses the method nor names the models. Two vendor evals are an announcement with numbers attached, not evidence. Verification waits until somebody outside can measure it.