Meta unveils new frontier-class AI models to compete with OpenAI, Anthropic

Driving the news: Meta on Monday unveiled Muse Wonder and Muse Wonder Full, models it says deliver state-of-the-art performance on a range of tasks involving personal and professional agents, memory and coding.

• The company posted a range of benchmarks showing the models, previously known by the code-name Watermelon, topping or matching the latest offerings from OpenAI and Anthropic.
• Muse Wonder is available through the Meta Model API and Muse Code, while the more capable Muse Wonder Full is available to Muse Code subscribers.

• Muse Wonder will also power future experiences in Meta's Muse assistant, initially for paid subscribers.
• The company says the models were trained using its gigawatt-scale Prometheus data center.

What they're saying: "These are true frontier models, competitive with the best models in the world, and are particularly state of the art on personal and professional agentic work," Meta Superintelligence Labs head Alexandr Wang tells Axios.

Zoom in: Meta is emphasizing two capabilities — "Active Memory" and "Active Focus" are designed to help AI systems perform more useful work over extended periods.

• Active Memory allows the AI to retain important information about a user across conversations and retrieve it when relevant to a new task.
• Active Focus is designed to keep an AI agent working toward a goal, including trying different approaches and adapting to new information when it encounters obstacles.
• Meta says the models also perform strongly on coding tasks and specialized professional work in areas including law, finance and health.
• "Memory is incredibly important for personal agents, and it means that it carries what matters about you across conversations over days, weeks, months, so that nobody is re-explaining themselves every time," Wang says.

Between the lines: Meta is positioning the new models as evidence that it can compete with the leading AI labs, rather than simply distribute AI technology through its enormous consumer platforms.

• The company says the models perform especially well on unfamiliar tasks, suggesting that their capabilities extend beyond the benchmarks and tasks emphasized during training.

The big picture: Meta's long-term ambition is what CEO Mark Zuckerberg has called "personal superintelligence" — AI that understands an individual's goals and preferences.

• Meta says the new models are a step toward that vision, with future specialized models intended for broader deployment across its products.

Reality check: Agents that remember personal information and act on a user's behalf also introduce new privacy and security challenges.

Between the lines: Meta isn't saying whether it will open source any of the Wonder family of models, but Wang noted that Meta recently open sourced the weights for Muse Glimmer as well as the technology needed for folks to put their Muse agent onto hardware.

• "We're still committed to open source," Wang said.

What to watch: Whether Meta's claimed performance advantages hold up in independent testing — and how quickly the new technology reaches the billions of people using its apps.

• On the safety front, Meta says it has strengthened safeguards against cyber and biological misuse, as well as protections for personal information. A model card with some details is set to be released on Monday, with a fuller report forthcoming.

Editor's note: This story has been corrected to say Meta's open-weight model is Muse Glimmer (not Glitter).