Meta introduces Muse Wonder and Muse Wonder Full, saying they deliver state-of-the-art performance on personal and professional agentic tasks (Meta AI Research)

Meta has introduced the Muse Wonder series, consisting of Muse Wonder and Muse Wonder Full, which showcase state-of-the-art performance on personal and professional agent tasks. These models represent the next step in Meta's scaling ladder and a significant leap toward personal superintelligence. Muse Wonder incorporates Active Memory and Active Focus to keep track of important information and work diligently toward goals.

Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is available for Muse Code subscribers. The Muse Wonder series will soon power experiences within Muse and power artifact generation and processes that require greater intelligence. The development of specialized models for broader use is planned for the future.

Meta's Muse Wonder series was developed after a fourteen-month overhaul of their AI efforts, moving from rebuilding their stack from first principles to pushing the frontier. The release of Muse Wonder and Muse Wonder Full builds on previous work in pretraining and reinforcement learning, resulting in the most capable models for complex tasks at work and in everyday life.

These models achieve state-of-the-art results across key benchmarks for personal and professional agents, memory, and coding when evaluated against other frontier models on various tasks. They demonstrate specialized knowledge applicable to work and daily life scenarios, such as understanding legal documents or processing financial and health information.

Active Memory and Active Focus are critical components of Muse Wonder and Muse Wonder Full, enabling them to remember essential context across conversations and continuously adapt to new information. These capabilities are essential for creating reliable personal and professional agents. The Muse Wonder series also excels in coding tasks, supporting projects from small applications to complex software.

The models were trained on diverse, long-running tasks to help them maintain progress throughout the software development process.