Source: Meta AI Research blog post of October 12, 2026 and the official Muse Wonder Model Card published by Meta Platforms Ireland Limited.
The architecture introduces two primary agentic mechanisms: Active Memory and Active Focus. Active Memory retains critical user context across multi-turn sessions ranging from 64,000 to 100,000 tokens while resisting injected distractor turns, scoring 85.5 percent on MemoryBench. Active Focus governs task persistence by attempting alternative pathways when blocked and requesting human intervention before destructive system actions.
In evaluations, Muse Wonder Full recorded 13.1 percent on Harvey's Legal Agent Benchmark (LAB-AA 1.1) and 59.1 percent on Vals Finance Agent V2, leading rival frontier systems including Claude Fable 5.1 (6.4 percent on Harvey LAB-AA) and GPT-6 Astra (8.6 percent). On SWE-Marathon, an ultra-long-horizon engineering benchmark from Abundant AI, Muse Wonder Full achieved 57.5 percent pass@1 with extra-high reasoning effort, compared to 42.5 percent for GPT-6 Astra.
On cyber capability tests, Muse Wonder Full captured 92 flags on ExploitGym, though an automated verification pass confirmed only 21 utilized the author-intended vulnerability while 71 exploited unintended flaws. On catastrophic safety assessments under the Meta Superintelligence Scaling Framework, the unmitigated model reached the high-risk threshold in chemical and biological domains before post-training and system-level classifiers lowered residual risk to moderate or lower.
Limits: Meta explicitly notes that 3D visual and spatial reasoning remains a known limitation, and the models are not licensed or intended for individuals under 18 years of age. Consumer pricing tiers for dedicated inference within the primary Muse application have not yet been announced.