Meta Unveils Muse Wonder Frontier Models for Autonomous Agents with Active Memory and Focus

Meta officially introduced Muse Wonder and Muse Wonder Full, its next-generation frontier reasoning foundation models engineered specifically to power autonomous personal and professional artificial intelligence agents. Trained across Meta gigawatt-scale Prometheus computing cluster, the 1-million-token context models introduce two breakthrough behavioral systems—Active Memory, which maintains persistent user context and preferences across weeks of interaction, and Active Focus, which enables agents to navigate obstacles and autonomously pursue complex multi-step objectives across real-world software environments.

Active Memory and Active Focus: The Dual Pillars of Agentic Continuity

Unlike conversational foundation models that reset their state after a session concludes, Muse Wonder is architected for continuous, long-horizon autonomy. Active Memory allows the model to retain critical personal preferences, workflow rules, and project context across disparate sessions without context decay, achieving a category-leading score of 85.6 on MemoryBench evaluations. In parallel, Active Focus equips agents with resilient task pursuit—enabling the model to diagnose execution roadblocks, adapt tool strategies dynamically, and seek human confirmation only before executing irreversible external actions.

Architecture & Performance Highlights

• Active Memory Integration: Retains multi-week operational context and user habits, outscoring rival frontier models on MemoryBench.
• Active Focus Resilience: Scored 54.2% on FocusBench for long-horizon task completion, surpassing GPT-6 Astra (47.7%) and Claude Opus 5.5 (45.8%).
• Long-Horizon Coding Superiority: Captured the top rank on SWE-Marathon (57.5%) and DeepSWE (79.0%) for multi-file terminal code refactoring.
• 1M Multimodal Context Window: Accepts multimodal inputs across text, video, audio, image, and PDF files with configurable reasoning effort levels.
• Dual Tier Distribution: Muse Wonder deploys via the Meta Model API ($4/M input, $20/M output), while Muse Wonder Full powers Muse Code workstations.

Prometheus Datacenter Training and Benchmark Leadership

Trained through extensive reinforcement learning scaled across thousands of synthetic failure environments, Muse Wonder is engineered to anticipate real-world software edge cases before they induce execution halts. In third-party evaluations, Muse Wonder established leading scores across 7 of 14 standard agentic benchmarks, demonstrating exceptional generalization to unfamiliar, out-of-distribution workflows. The release serves as the intelligence engine for Meta broader consumer hardware and developer tools, including Muse Code, Ray-Ban Meta glasses, and open-source Muse Gadgets.

Model Profile Sheet