Use Meta's Muse Wonder for complex tasks with active memory and self-correction features

Meta Platforms Inc. has launched two new artificial intelligence models, Muse Wonder and Muse Wonder Full, which the company describes as its most capable agents for handling complex tasks. According to a report by SiliconANGLE, these models are targeted at longer, multistep jobs that users typically delegate to AI assistants, such as navigating an insurance claim or performing specialized work in law, finance, and healthcare.

In plain English

These models introduce two primary features to improve user experience. Active Memory allows the AI to carry context from previous conversations into current ones, surfacing relevant details when they become important. Active Focus is designed to help the model persist through difficult tasks. When the model encounters an error or gets stuck, it attempts alternative approaches and checks with the user before executing significant actions, rather than failing silently or proceeding incorrectly.

The background

Meta trained these models at its Prometheus data center, which operates at gigawatt scale. The development builds on pretraining and reinforcement learning techniques previously introduced with the Muse Spark model in April. Users can input prompts that combine text, video, audio, and images within a one-million-token context window, though the model’s responses remain text-only. This multimodal input capability allows for more nuanced task definitions, although the output format remains constrained to written text.

What changes now

Meta compared Muse Wonder Full against competitors including OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 and Claude Opus 5.5. In legal benchmarks, Muse Wonder Full achieved a 13.1% pass rate, outperforming GPT-6 Astra’s 8.6%. Coding results were mixed; Muse Wonder Full led in long-horizon tests like SWE-Marathon but trailed Claude Opus 5.5 in ProgramBench. Claude Opus 5.5 also outperformed Meta’s model in health benchmarks and general computer-use tests.

What to do and how to stay safe: Meta

• Evaluate whether your organization’s workflows involve complex, multistep tasks that could benefit from agent-based automation, particularly in legal or coding contexts.
• Review the model’s error-checking behaviors, such as Active Focus, to determine if the human-in-the-loop verification steps align with your security and compliance requirements.
• Test the multimodal input capabilities with non-sensitive data to assess how well the model handles mixed media prompts within the one-million-token context window.
• Monitor benchmark comparisons provided by the vendor to understand relative strengths and weaknesses against competing models like GPT-6 Astra and Claude Opus 5.5 before deploying in production.

Step-by-step guide: Stop AI Chatbots Leaking Sensitive Business Data

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

Frequently asked questions

What new features distinguish Muse Wonder from previous Meta AI models?

Muse Wonder introduces Active Memory, which retains context across conversations, and Active Focus, which helps the model recover from errors and persist through difficult tasks.

How does Muse Wonder Full perform compared to competitors in legal and coding tasks?

Muse Wonder Full outperformed GPT-6 Astra in legal benchmarks with a 13.1% pass rate and led in some coding tests, though it trailed Claude Opus 5.5 in others like ProgramBench.

Can users input images or video when using Muse Wonder?

Yes, prompts can mix text with video, audio, and images within a one-million-token context window, although the model’s responses are returned only in text format.