Meta says new muse model hits AI frontier after aggressive push

Meta Platforms Inc. unveiled a new artificial intelligence model on Monday called Muse Wonder, a milestone in the company's multi-billion dollar push to revamp its AI organization. This development comes as Meta's shares surged by up to 4.7% following the announcement, with the stock gaining nearly 9% so far this year. The company's new model, known internally as "Watermelon," represents a significant leap over its previous AI iterations, boasting "frontier capabilities" that its leadership claims are on par with those of OpenAI's ChatGPT and Anthropic PBC's Claude when it comes to coding, problem-solving, and working through complex questions.

While comparing AI models can be challenging, with some excelling in certain areas and others lagging, if Meta's claims prove accurate and Muse Wonder gains traction among researchers and developers, it would mark a substantial turnaround for the company. Last year, Meta's AI models struggled to make an impact in Silicon Valley.

Meta AI Chief Alexandr Wang described Muse Wonder as a "true frontier model" and a competitive contender with the best models available today, particularly in personal and professional tasks. Wang emphasized that the company aimed to differentiate itself by focusing on broad consumer uses for these large models. The Muse Wonder model will soon power the subscription version of Meta's personal AI agent, designed to assist users with everyday tasks and errands.

One of the key features of Muse Wonder is its ability to maintain "active memory" and "active focus," enabling the AI agent to remember relevant information about a user over extended periods, even as tasks become increasingly challenging. This means users won't have to repeatedly explain themselves to the AI agent. Wang and his Meta Superintelligence Labs team are already working on developing even larger and more scaled models, while also striving to reduce the cost of serving this technology to billions of users.

Muse Wonder is available to developers through the Meta Model API and Muse Code. Wang encouraged skeptics who remain unconvinced that Meta has reached the AI frontier to try the model and make their own judgments. He expressed confidence in the model's performance and dismissed doubts from critics.