Meta has launched Muse Wonder and Muse Wonder Full, two artificial-intelligence models it says can compete with the leading systems from OpenAI and Anthropic. The release is an important test of Meta's heavy spending on AI researchers and computing infrastructure.
The company is pitching Wonder as more than another chatbot upgrade. Meta says the models are designed for longer-running personal and professional tasks, including coding and specialized work in law, finance and health. They were previously known by the code name Watermelon and were trained using Meta's gigawatt-scale Prometheus data center, Axios reported.
What Meta says the models can do
Meta is emphasizing two features called Active Memory and Active Focus. Active Memory is meant to retain relevant information about a user across conversations, while Active Focus is intended to keep an agent working toward a goal, trying different approaches when it hits an obstacle.
That combination points toward Meta's larger ambition of building a personalized assistant that can remember a user's preferences and act on their behalf. The company plans to use Muse Wonder in future versions of its Muse assistant, beginning with paid subscribers.
Meta has also published internal benchmarks that it says show the Wonder models matching or beating recent offerings from OpenAI and Anthropic. Those results remain company claims, not independent validation. The practical test will come as outside researchers and customers evaluate the models on their own work.
Access starts with developers and subscribers
Muse Wonder is available through the Meta Model API and Muse Code. Meta describes Muse Wonder Full as the more capable model and is limiting it to Muse Code subscribers for now.
The launch also drew an immediate response from investors. MarketWatch reported that Meta shares were up more than 3% on Oct. 12 after the announcement, as the company presented the release as a return to the top tier of AI development.
The unresolved questions
Meta has not said whether it will release the Wonder family as open-weight models. Axios noted that Meta recently open-sourced the weights for Muse Glimmer and technology for putting a Muse agent onto hardware, and Meta Superintelligence Labs head Alexandr Wang said the company remained committed to open source.
The features that make the models attractive also raise familiar risks. A system that remembers personal details and acts over long periods can be more useful, but it also creates privacy and security concerns. Meta says it has strengthened protections against cyber and biological misuse and added safeguards for personal information. Axios reported that a model card with some details was set for release Monday, with a fuller report still to come.
For Meta, the release is a bid to show that its expensive AI reset has produced frontier-level technology. For users, the more consequential questions are how well the models perform outside Meta's own tests, what data persistent memory retains, and how quickly the technology moves from developer tools into products used by billions of people.
