Meta has released Muse Wonder, a family of autonomous language models designed to retain user context over long periods. The systems incorporate active memory tools to complete complex professional workflows without constant user prompting. The base version is available now through the Meta Model API and Muse Code, while the advanced edition requires a paid subscription.
The release includes 2 distinct models. The standard Muse Wonder model provides core processing for general coding and research, while the larger Muse Wonder Full targets demanding enterprise work. Meta opened access to the standard build through its Model API and the Muse Code programming interface. Access to Muse Wonder Full remains restricted to paying Muse Code subscribers, with broader rollout to consumer services planned for later stages.
The system relies on two primary mechanisms. An engine called Active Memory stores details about individual user habits and project histories, pulling relevant data when needed for specific calculations. A complementary tool named Active Focus manages long running tasks. Instead of stopping when it hits errors, the model tests alternative strategies and processes new input independently. It only pauses execution when it requires human confirmation for critical actions.
Meta trained both models on specialized datasets covering legal, medical, and financial fields. Training focused heavily on error correction and multi step verification, allowing the software to handle unfamiliar problems better than older systems. Because the models can read email inboxes, calendar entries, and internal documents, Meta implemented system protections to stop data leakage and block malicious user prompts.
