AI
Meta deploys Wonder-trained AI models across Muse and Meta AI for all users
Meta is rolling out new specialized models trained on its frontier Wonder AI to power its consumer agents, adding memory and task-persistence features without requiring users to access the expensive core model.
1/3
The New Stack
Starting Friday, users of Meta's Muse agent and Meta AI gain access to new AI models that Meta trained using Muse Wonder, its most advanced model. While premium subscribers retain direct access to Wonder for complex tasks, the general user base now benefits from Wonder's capabilities through these specialized derivatives.
The new models lack official names from Meta, but the company emphasizes they are purpose-built versions of existing models that underwent training with Wonder as a guide. Both come equipped with Active Memory and Active Focus—two capabilities Meta first introduced alongside Wonder itself.
Trickle-down AI at consumer scale#
Meta launched Wonder and Wonder Full in mid-October, initially restricting access to the company's Model API and Muse Code (its coding assistant). The full-powered variant remains available only to the highest-tier Muse Code subscribers.
Muse, Meta's personal agent application that debuted in September, has become a notable success story for the company. Built on the smaller Muse Spark model, it attracted significant consumer interest by offering a free tier—a decisive advantage over competing agents like OpenAI's Dots or Anthropic's Duet, both of which require premium subscriptions.
Rather than simply scaling Wonder down into a general-purpose model, Meta took a different approach: the company trained these new models specifically on the behaviors and decision patterns that matter in a consumer product. The agents learned when to operate autonomously versus when to seek user confirmation, and how to recover from mid-task corrections without losing context.
This knowledge-transfer strategy is commonplace among AI labs seeking to distribute frontier capabilities across production systems, though Meta has not disclosed its specific methodology.
Why specialized beats general#
A personal agent needs breadth—booking restaurants, managing inboxes—but the underlying model doesn't require the same comprehensive knowledge as a frontier model. What it does need is the ability to recall details from conversations days earlier and gracefully handle website errors mid-transaction.
That's where Active Memory and Active Focus provide the difference.
Active Memory enables models to distinguish between temporary and permanent information across conversations, reason about time, and identify what remains relevant versus what has grown stale. In Meta's internal testing, the new Muse models achieved 90% accuracy in applying critical context, up from 76% with the previous generation.

