Meta releases Muse Wonder models for personal and professional agents

Meta released Muse Wonder and Muse Wonder Full, which it says are its most capable models for running multistep agent tasks. Muse Wonder Full passed 13.1% of tasks on Artificial Analysis' version of Harvey's Legal Agent Benchmark, ahead of GPT-6 Astra at 8.6%. Meta's model card notes that its competitor runs were best-effort and that the SWE-Marathon comparison is not controlled.