Meta on October 12 introduced two new models designed for agentic tasks, Muse Wonder and Muse Wonder Full, adding cross-conversation memory and sustained task handling to its foundation-model lineup.
Immediate access depends on developer deployment paths. Meta has released Muse Wonder through the Meta Model API and within its Muse Code developer platform. Muse Wonder Full is available to Muse Code subscribers. Meta says Wonder-powered experiences in its consumer Muse product are coming soon, with subscribers first. It also plans specialized models for broader Muse access.
Memory, Focus and Coding
Building on the company's prior Muse Spark foundation, Meta describes two features of the new models as Active Memory and Active Focus:
- Active Memory: Designed for cross-conversation context retrieval, allowing an agent to retrieve relevant details from earlier interactions.
- Active Focus: Meta’s term for sustained problem-solving. According to the company, this allows agents to explore alternative approaches during multi-step tasks, adapt dynamically as new information emerges, and ask for clarification and check in before important actions.
For coding implementations within Muse Code, Meta describes these capabilities as tracking multi-part technical specifications and running repeated test-and-revision cycles without losing project context.
Infrastructure Claims and Pending Safety Data
Meta says the models build on larger, stronger base models and were trained at its gigawatt-scale Prometheus data center. In its announcement materials, the company asserts both evaluation superiority and safety improvements compared to prior iterations.
The evaluation comparisons are Meta-reported results. The company says a fuller report with additional safety evaluations will follow.