Muse Wonder

Muse Wonder is a new model series - Muse Wonder and Muse Wonder Full - designed to power personal and professional agents by combining two core innovations: Active Memory, which carries and retrieves user-relevant context across interactions, and Active Focus, which sustains multi-step work by trying alternatives, resolving blockers, and asking clarifying questions. Trained at Meta’s gigawatt-scale Prometheus data center and built on prior scaling work from Muse Spark, the models achieve state-of-the-art results across agent, memory, and coding benchmarks and demonstrate domain expertise in law, finance and health. Real-world demos show them building a playable 3D kitchen game from images, planning a personalized run, arranging complex wedding seating, and designing custom swag from social photos. Models are available via the Meta Model API and Muse Code, with Muse Wonder Full offered to Muse Code subscribers, and will power Muse experiences and artifact pipelines.

A key finding is improved generalization: Muse Wonder Full produces stronger out-of-distribution performance relative to its in-distribution gains, attributable to a larger, stronger base model and post-training methods. Evaluation suites include Harvey’s Legal Agent Benchmark, FocusBench and SWE‑Marathon for sustained execution, and frontier coding tests that train models on long-running software tasks. Safety and alignment receive heavy emphasis: training targets robustness to adversarial instructions, discretion with private data, and strong safeguards against CBRN and cyber misuse while preserving defensive cyber capabilities under Meta’s Superintelligence Scaling Framework. The launch presents a major technical step toward personal superintelligence, prioritizing sustained work, memory, domain skill, and safety in deployed agents.