Muse Wonder(research.meta.ai)
662 points by meetpateltech 1 day ago | 375 comments
tl;dr: Meta introduces Muse Wonder and Muse Wonder Full, frontier models aimed at personal and professional agentic tasks, available via the Meta Model API and Muse Code. The models emphasize "Active Memory" (retaining relevant user context) and "Active Focus" (sustained goal pursuit), claiming state-of-the-art results on benchmarks in domains like law, finance, health, and coding (including SWE-Marathon). Meta also highlights improved out-of-distribution generalization from stronger base models, and stronger safeguards against CBRN and cyber misuse under its Superintelligence Scaling Framework.
HN Discussion:
  • ↑More frontier model competition prevents dangerous concentration of AI power
  • ↓Security and privacy concerns with Meta exposing user profiles on login
  • ↓Real-world testing shows model underperforms competitors like Opus 5.5 despite benchmark claims
  • ↓Meta failed to deliver on previous open-weights promises for Muse Spark
  • ↑Strong benchmark performance suggests frontier models are becoming commoditized