Muse Wonder

Meta has unveiled the Muse Wonder series, its most capable models yet, designed to excel at personal and professional agentic tasks. The Muse Wonder model and its enhanced version, Muse Wonder Full, both mark a significant leap forward in AI development. These models integrate Active Memory and Active Focus, enabling them to keep track of important information and work diligently towards goals.

Muse Wonder is accessible now via the Meta Model API and Muse Code, while Muse Wonder Full is available for Muse Code subscribers. Muse Wonder will soon power experiences within Meta's Muse platform, starting with subscribers. Beyond powering experiences and artifact generation, Muse Wonder will also play a crucial role in developing specialized models tailored for Meta's user base.

The Muse Wonder series is a result of a fourteen-month overhaul of Meta's AI efforts, pushing the frontier of what's possible in artificial intelligence. Trained at Meta's gigawatt-scale Prometheus data center, these models demonstrate state-of-the-art performance across key benchmarks for personal and professional agents, memory, and coding.

They achieve this by leveraging specialized knowledge to handle complex tasks in areas such as law, finance, and health. Muse Wonder and Muse Wonder Full both demonstrate superior performance across these benchmarks compared to other frontier models. A closer look at four areas central to their work reveals that these models excel in domain expertise, Active Memory, Active Focus, and frontier coding.

Active Memory allows the models to apply specialized knowledge to real-world problems, while Active Focus enables them to remember what's important and retrieve relevant details during tasks. Sustained effort and adaptability are key to Active Focus, which the models demonstrate through FocusBench and SWE-Marathon evaluations. In terms of coding, the models showcase frontier coding performance, capable of handling projects from small applications to complex software.

Despite similar benchmark scores, one model may exhibit a broader range of capability, a phenomenon often referred to as "benchmaxxing" or "big model smell." To quantify this, the Muse Wonder Full model outperforms Muse Spark 1.3 on unfamiliar tasks when scaled with test-time compute, thanks to its stronger base model. These models are equipped with robustness to adversarial instructions and careful handling of private information, ensuring safe and responsible use.

As personal agents become more capable, Meta's goal is to help people harness these capabilities while maintaining privacy and security.