Muse Wonder and Muse Wonder Full: Meta claims to have developed a state-of-the-art model for personal and professional assistants

Trained on the Prometheus data centre, the two models rely on active memory and greater resilience in the face of obstacles. They are available exclusively to paying subscribers and developers.

On October 12, Meta introduced the Muse Wonder series, consisting of two models: Muse Wonder, accessible via the Meta Model API and the Muse Code code agent, and Muse Wonder Full, a more powerful variant reserved for Muse Code subscribers ($50 or $200 per month packages). The group claims that they reach the state of the art on personal and professional agent tasks, memory and programming. They were trained in Prometheus, Meta's gigawatt-scale data center, fourteen months after the complete overhaul of its AI teams.

Active memory and perseverance

Meta highlights two capabilities. "Active Memory" is about keeping what matters to the user from one conversation to the next and bringing it out at the right time. "Active Focus" refers to the ability to pursue a difficult goal by trying other avenues rather than stalling, while asking for confirmation before important actions. According to Alexandr Wang, head of AI at Meta, the model, codenamed "Watermelon", was trained precisely to bypass the unforeseen events of a real task such as a restaurant reservation.

The model will soon power the Muse app, initially for subscribers ($20 or $100 per month). Meta also plans to use it to develop specialized models distributed for free. Wang assures that access via the API will be charged significantly less than the cutting-edge models of OpenAI and Anthropic, while acknowledging that Meta monitors incidents related to the Muse agent, which some users have accused of having viewed private messages without authorization, which the company disputes.