Meta unveils frontier-class AI models in bid to catch OpenAI and Anthropic
Photo source: Brendan Lynch/Axios
The launch is a test of Meta’s heavy investment in AI research and computing infrastructure. The company is highlighting memory and agents that can work toward goals over extended periods.
Meta unveiled Muse Wonder and Muse Wonder Full on Monday, saying the models deliver state-of-the-art performance across tasks involving personal and professional agents, memory and coding. The company said the models, previously known by the codename Watermelon, top or match the latest offerings from OpenAI and Anthropic in a range of benchmarks.
This is reported by Axios.
Muse Wonder is available through the Meta Model API and Muse Code. The more capable Muse Wonder Full is available to Muse Code subscribers. Meta said Muse Wonder will also power future features in its Muse assistant, initially for paid subscribers. The models were trained using the company’s gigawatt-scale Prometheus data center.
Meta Superintelligence Labs head Alexandr Wang told Axios the models are designed for work involving personal and professional agents.
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,
– Alexandr Wang, head of Meta Superintelligence Labs
Meta is emphasizing two features: “Active Memory,” which can retain information about a user across conversations and retrieve it for a relevant task, and “Active Focus,” designed to keep an AI agent working toward a goal, try different approaches and adapt when it encounters obstacles. The company also said the models perform strongly on coding and specialized work in 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,
– Alexandr Wang
Meta said the models perform especially well on unfamiliar tasks, which it presents as evidence that their capabilities extend beyond benchmarks and tasks emphasized during training. The company is positioning them as proof it can compete with leading AI labs, rather than simply distribute AI through its consumer platforms.
The models are a step toward what CEO Mark Zuckerberg has called “personal superintelligence” – AI that understands an individual’s goals and preferences. Meta said it plans to develop specialized models for broader use across its products.
Agents that retain personal information and act on a user’s behalf also raise privacy and security challenges. Meta has not said whether it will open-source any Muse Wonder models. Wang noted that the company recently open-sourced the weights for Muse Glitter and technology that allows people to put their Muse agent on hardware.
We’re still committed to open source,
– Alexandr Wang
Questions ahead include whether the models’ claimed advantages hold up in independent testing and how quickly the technology reaches the billions of people who use Meta’s apps. Meta said it strengthened safeguards against cyber and biological misuse and added protections for personal information. A model card with some details was set to be released on Monday, with a fuller report to follow.
You may be interested in these materials:
- Meta is exploring a potential two‑year agreement to rent compute capacity from Anthropic worth up to $10 billion, aiming to diversify revenue beyond advertising and expand cloud AI offerings.
- Mistral has introduced ML4, an open-weight AI model the French company says can rival leading closed systems. The model’s performance claims have not been independently verified.
- Nvidia plans to acquire Hugging Face for $12.9 billion, strengthening its position in open-source AI and expanding access to models, datasets, and developer tools.