Meta Platforms Inc. has launched Muse Wonder and Muse Wonder Full, two AI models designed for personal and professional agent tasks. The models feature Active Memory for context retention and Active Focus for error recovery and task persistence.
The models support a 1 million-token context window and accept multimodal prompts including text, video, audio, and images. They were trained at Meta's gigawatt-scale Prometheus data center, building on methods introduced with Muse Spark in April.
In benchmarking, Muse Wonder Full achieved a 13.1% success rate on the Harvey Legal Agent Benchmark, compared to 8.6% for GPT-6 Astra. The model outperformed Claude Opus 5.5 in the SWE-Marathon coding test with a 57.5% score.
Meta reports that Muse Wonder Full shows a 0% harmful-action rate in agentic misalignment tests. The company plans to release a full safety report, noting that current safeguards keep risks at a moderate level or lower.
Muse Wonder replaces Muse Spark 1.3, which was released six weeks ago. The Muse personal agent has been powered by Muse Spark since its U.S. launch on September 8.
