Muse Wonder

Summary

Meta's announcement of Muse Wonder and the larger Muse Wonder Full, which it calls its most capable models for personal and professional agent tasks. They follow Muse Spark 1.1 (July 2026) and, like it, are offered only through Meta's API and products, not as open weights. The post names benchmarks but gives no scores, and promises a separate results report and a safety report.

Key points

• Availability (2026-10-12): Muse Wonder through the Meta Model API and Muse Code; Muse Wonder Full only for Muse Code subscribers. It will later power the Muse app, starting with subscribers. No prices were given.
• Active Memory: carries context about the user across conversations and pulls in relevant details for the current task.
• Active Focus: keeps working on hard goals by trying other approaches, working around blockers, and checking in before important actions.
• Claims state-of-the-art results on agent, memory, and coding benchmarks, naming Harvey's Legal Agent Benchmark (LAB), FocusBench, and SWE-Marathon. No numbers in the post itself.
• Trained at Meta's Prometheus data centre on "larger, stronger base models"; no parameter counts. Meta calls it its "most aligned model to date", with stronger CBRN and cyber safeguards.
• No mention of open weights or a licence.

Why it matters

Meta's frontier line remains closed and API-first, and it now competes directly on long-running agents with built-in cross-session memory. The benchmark claims cannot be checked until Meta publishes the results report.