Meta's Muse Wonder AI aims to streamline professional workflows with enhanced error recovery and active memory

Meta Platforms Inc. released Muse Wonder and Muse Wonder Full today, describing them as its most capable models for running agents that manage personal and professional duties. According to SiliconANGLE, the company targeted these releases at longer, multistep jobs that users typically delegate to AI assistants, such as navigating insurance claims.

What new capabilities do these models offer?

The models introduce two primary features to improve agent performance. Active Memory allows the system to carry user context from one conversation to the next, surfacing relevant details when they become important. Active Focus is designed to help the model persist through difficult tasks.

According to Meta’s announcement, a model stuck on a task will attempt alternative approaches and recover from errors. It also checks in with the user before executing important actions. The system accepts prompts mixing text, video, audio, and images within a 1 million-token context window, though responses are returned only in text format.

How do they perform in benchmarks?

Meta’s model card compares Muse Wonder Full against OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 and Claude Opus 5.5. In Artificial Analysis’ version of Harvey’s Legal Agent Benchmark, Muse Wonder Full passed 13.1% of tasks, outperforming GPT-6 Astra’s 8.6%.

Coding results were split. Muse Wonder Full led DeepSWE 1.1 and SWE-Marathon, scoring 57.5% on the latter versus 46.3% for Claude Opus 5.5. However, the card notes these are not controlled comparisons. Claude Opus 5.5 won ProgramBench with 70.5% to 59%, and also led HealthBench Pro and OSWorld V2.1.

What is the technical foundation?

Meta trained both models at its gigawatt-scale Prometheus data center. The training builds on pretraining and reinforcement learning methods introduced with Muse Spark in April. SiliconANGLE reported that Meta aimed the release specifically at complex agent work rather than simple queries.

The company stated that coding remains a strength, along with specialist fields such as law, finance, and health. The model card cautions that Meta’s runs of competing models were best-effort and that prompts may not have been optimally tuned for those rivals.

What are the implications for developers?

Developers can expect these models to handle more autonomous, long-horizon tasks than previous iterations. The inclusion of Active Memory suggests a shift toward persistent agent interactions that remember user preferences across sessions. However, the mixed benchmark results indicate that performance varies significantly by task type.

SiliconANGLE noted that Meta positioned these models for professional and personal agents. The ability to recover from errors via Active Focus may reduce the need for constant human oversight in complex workflows, though the company emphasizes user confirmation for critical steps.

What to do and how to stay safe: Meta

• Review the model card for specific benchmark limitations and reasoning settings before integration.
• Test the models with your specific multistep workflows to verify error recovery behavior.
• Monitor the Active Memory feature for data retention compliance in your organization.
• Compare performance against current tools using your own internal metrics rather than relying solely on vendor benchmarks.

Step-by-step guide: AI Usage Policy Checklist for Secure Enterprise Deployment

General security guidance from the Sandbox Desk newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What are the names of the new models?

Meta released Muse Wonder and Muse Wonder Full.

Do these models support multimodal inputs?

Yes, they accept text, video, audio, and images within a 1 million-token context window.

Which model performed best in coding benchmarks?

Results were mixed; Muse Wonder Full led SWE-Marathon, while Claude Opus 5.5 led ProgramBench.