Meta releases Muse Wonder AI models for complex agent tasks
Meta launches Muse Wonder and Muse Wonder Full, claiming they outperform rivals in legal benchmarks and long-horizon coding tests.
Key points
- Meta released Muse Wonder and Muse Wonder Full, models designed for multistep personal and professional agent tasks.
- The models feature Active Memory for context retention and Active Focus for error recovery during difficult tasks.
- Benchmarks show Muse Wonder Full leading in legal agent tests and specific coding challenges, though rivals lead in others.
Meta Platforms Inc. released Muse Wonder and Muse Wonder Full today, introducing two artificial intelligence models it claims are its most capable for handling extended, multistep agent workflows. The company positioned these releases for complex personal and professional jobs, such as navigating insurance claims or managing specialist tasks in law, finance, and health sectors.
These models incorporate two new capabilities: Active Memory, which retains user context across conversations, and Active Focus, which allows the system to attempt alternative approaches when stuck. Meta stated that Active Focus enables the model to recover from errors and check with users before executing significant actions, aiming to improve reliability in long-running tasks.
Technical Architecture and Training
Meta trained both models at its Prometheus data center, utilizing pretraining and reinforcement learning techniques first introduced with Muse Spark in April. The system supports a one-million-token context window, allowing prompts to combine text, video, audio, and images. However, the models currently return answers only in text format, maintaining a unimodal output despite multimodal input capabilities.
Performance Benchmarks
According to the model card published by Meta, Muse Wonder Full achieved a passing rate of 13.1% on Artificial Analysis’ Harvey’s Legal Agent Benchmark, surpassing GPT-6 Astra’s 8.6%. In coding, Muse Wonder Full led DeepSWE 1.1 and SWE-Marathon, scoring 57.5% on the latter compared to Claude Opus 5.5’s 46.3%. Meta noted these coding comparisons were not controlled, as rival scores used different tools and reasoning settings.
Competitors remained stronger in other areas. Claude Opus 5.5 outperformed Muse Wonder Full on ProgramBench, HealthBench Pro, and the OSWorld V2.1 computer-use benchmark. Meta cautioned that its evaluations of OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 and Opus 5.5 were best-effort runs, with prompts potentially not optimized for those specific models.
What should you do?
Security and technology professionals should monitor official model cards for benchmark limitations when evaluating new AI agents. Teams should test multimodal input handling in isolated environments before deploying agents that access sensitive legal or financial data. Verify that error-recovery mechanisms like Active Focus do not bypass existing approval workflows for critical actions.
What to do and how to stay safe: Meta
- Review vendor-provided model cards to understand benchmark limitations and testing conditions before integration.
- Test new AI agents in sandboxed environments to verify behavior during error recovery and context retention.
- Ensure that any automated actions taken by the agent require human approval for high-impact decisions.
- Monitor for updates on multimodal input security, as text-only outputs may still reflect risks from complex inputs.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What new features do the Muse Wonder models offer?
They include Active Memory for carrying context across conversations and Active Focus for attempting alternative approaches when tasks become difficult.
How did Muse Wonder Full perform in legal benchmarks?
It passed 13.1% of tasks in the Harvey’s Legal Agent Benchmark, according to Meta’s model card comparison.
Which rival models performed better than Muse Wonder Full?
Claude Opus 5.5 scored higher on ProgramBench, HealthBench Pro, and the OSWorld V2.1 computer-use benchmark.
