Meta launches Muse Wonder, its first frontier model

Summary

Meta has launched Muse Wonder and Muse Wonder Full, a new family of frontier models for personal and professional agents. The company presents them as its most capable models so far, with a focus on coding, specialised knowledge work, and tasks that require multiple steps.

In practice

Muse Wonder is available through the Meta Model API and Muse Code; Muse Wonder Full is currently restricted to Muse Code subscribers. According to the model card, the models take text, images, video, and audio as inputs, generate text, and support up to one million tokens of context.

Meta highlights two capabilities: “Active Memory”, intended to retain and retrieve relevant information across conversations, and “Active Focus”, intended to sustain work on a complex goal, adapt to new information, and ask for clarification before important actions.

What we still don't know

Many of Meta’s published performance comparisons come from the company itself. The model card combines internal tests, provider-published scores, and third-party evaluations, while acknowledging that configuration, tools, and prompting can affect results.

Meta also identifies limitations: the model can produce inaccurate or biased answers, make errors in multi-step reasoning, and vary in performance across languages, serving configurations, and tools. The company says a fuller safety report will be released later.

Why it matters

• Meta is moving from lighter models to an offering intended to compete in agentic work and coding.
• Memory and sustained execution only become useful with permission controls, human confirmation, and evaluation in the real use case.
• Wonder Full’s limited availability makes it too early to assess its impact outside the Muse ecosystem.