On October 12, Meta unveiled Muse Wonder and Muse Wonder Full — its most powerful artificial intelligence models. According to the company, they deliver the best results in tasks for personal and professional AI agents, memory handling, and programming.
Muse Wonder is already available via the Meta Model API and the Muse Code service, while the advanced Muse Wonder Full version is available to Muse Code subscribers. Later, the model will also be integrated into consumer Muse products: initially for paying users, as well as for generating artifacts and handling more complex processes.
The models were trained in Meta's gigawatt-scale data center, Prometheus. According to Axios, they were internally codenamed Watermelon. Benchmarks published by Meta show that the new models outperform or match the latest developments from OpenAI and Anthropic.
Meta highlights two key innovations: Active Memory and Active Focus. The former enables the model to remember important user context across conversations and retrieve relevant details during task execution. The latter is responsible for sustained work on complex goals: the model tries alternative approaches, navigates obstacles, clarifies details, and double-checks before taking critical actions.
The company also emphasizes programming and domain expertise. The models were tested on long-running software projects (SWE-Marathon benchmark) as well as tasks in law, finance, and medicine — including the legal benchmark Harvey LAB.
“These are truly cutting-edge models that compete with the best in the world,” Meta Superintelligence Labs head Alexander Wang told Axios. The release is the result of a 14-month overhaul of Meta’s AI division, which began with rewriting its technology stack from scratch.
Regarding safety, Meta claims that Muse Wonder features the strongest abuse prevention in its history across chemical, biological, radiological, and nuclear weapons, as well as cyberattacks, and handles private data more cautiously. The company promises to publish a full safety report soon. It remains to be seen how much independent testing will confirm these claimed advantages.
Sources: Meta AI Research, Axios.