Meta Launches Muse Wonder, Its Most Capable Model Yet, Built for AI Agents

Meta is launching Muse Wonder, a new model built for AI agents that handle personal and professional tasks on your behalf.

The company announced it on October 12, 2026, at 17:08 UTC, alongside a more powerful version called Muse Wonder Full. Both build on the groundwork Meta laid with Muse Spark.

Muse Wonder for developers, Wonder Full for Muse Code subscribers

Access differs between the two models. Here is how it looks, according to the official announcement:

Model	Available today	Coming next
Muse Wonder	Meta Model API and Muse Code	Will power Muse experiences, subscribers first
Muse Wonder Full	Muse Code subscribers	Not announced

Meta says the model will also power artifact generation and processes that need more intelligence. Later, the company plans to use it to develop specialized models for everyone on Muse.

Memory that carries over, and a model that keeps going

Meta pitches both models as having expertise in law, finance and health. Its examples: navigating an insurance claim, understanding a lease and making sense of personal finances.

Active Memory is the ability to carry what matters to you across conversations and pull up the details relevant to the task at hand.

Active Focus is meant for hard tasks. The model tries other approaches, recovers from errors and follows through. Meta says it also asks for clarification when needed and checks in before anything important.

For coding, Meta claims frontier performance, from small apps to complex software. In one demo, the model builds a playable 3D kitchen game.

A 1M-token context, with image, video and audio input

According to the model card, Muse Wonder takes text, images, video and audio as input and returns text. Its context window reaches 1 million tokens.

Meta trained the models at its gigawatt-scale Prometheus data center. It describes them as the next step on its scaling ladder, across pretraining and reinforcement learning.

The model card also states that the model is not intended for use by anyone under 18.

Safety: privacy, refusing harm and cyber defense

A personal agent works across your messages, calendar and documents. So Meta trained the models to avoid sharing private information without permission and to turn down requests to help with actually harmful acts.

Security teams can still use the model to find flaws in their own code. Meta says it follows the Meta Superintelligence Scaling Framework, and a full safety report is coming soon.