Meta released Muse Wonder and Muse Wonder Full. Both models accept text, image, video, and audio inputs. They target commercial and research use in agentic tasks. The context length is 1M tokens at inference. Evaluations cover personal and professional agents, coding, and general intelligence. Competitor comparisons include GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5. Risk assessments flagged Chemical & Biological as high risk without mitigations. Mitigations bring all domains to moderate or lower risk. Cyber capability scores include 88.4% on CyberGym and 58.9% on ExploitBench. Adversarial refusal rates are below 2.3% across benchmarks. Dual-use bio and chem knowledge scores average 60.7%. BioTIER refusal rates exceed 98% in English and multilingual tests. Loss of control evaluations show 87.3% honesty under pressure and 0% harmful action rate.

• 1M token context length
• 92 flags captured on ExploitGym
• 21 intended-vulnerability exploitations on ExploitGym
• 98.7% BioTIER refusal rate