Meta Says New Muse Model Reaches AI Frontier After Aggressive Push

Meta Platforms Inc. released a long-awaited artificial intelligence model on Monday, a milestone that will test whether its multibillion-dollar push to rebuild its AI organization has improved its technology enough to compete among leaders.

Known as Muse Wonder, the new model is the company’s most capable since Meta Chief Executive Officer Mark Zuckerberg overhauled his AI operation through an aggressive hiring push that put new talent at the helm. The company claims the model, dubbed “Watermelon” internally, is a big leap over its own predecessors, with “frontier capabilities” that its leadership say put it on par with OpenAI’s ChatGPT and Anthropic PBC’s Claude when it comes to coding, or working through financial or legal questions.

It can be difficult to compare AI models, as some are better at some tasks and worse at others, and benchmarking parameters can be gamed and aren’t necessarily reflective of how the technology performs in the real world. But, if Meta’s claims hold up to outside testing, and Muse Wonder gains traction with researchers and developers, it would mark a significant turnaround for the company, whose models last year fell flat in Silicon Valley.

“These are true frontier models,” said Meta AI Chief Alexandr Wang in an interview on Monday. “We really think they are competitive with the best models out there, and particularly state of the art on personal and professional agentic work.”

Meta’s shares rose as high as 4.7% on the announcement. The stock has gained nearly 9% so far this year through Friday’s close.

Over the last year and a half, Zuckerberg refocused the company’s efforts on advancing in the AI race, including by committing hundreds of billions of dollars to build out the infrastructure needed to support the models, and recruiting Wang to lead the new AI unit as part of a $14 billion investment into Scale AI .

For Wang, Muse Wonder’s release represents the culmination of that effort. “This is the big moment,” he said. Wang added that Meta has tried to differentiate itself from competitors by focusing on big models for broad, consumer uses. Last month, the company unveiled a personal AI agent known as Muse, which is designed to help with everyday tasks and errands. The new model will soon power the subscription version of that app.

“I think the rest of the ecosystem has been really focused on coding agents,” Wang said. “We do genuinely think that we’ve taken a differentiated focus to make the personal-agent use case really sing.”

Part of that effort has been to create so-called active memory and active focus capabilities that allow an AI agent to remember relevant information about a user over time and keep working when tasks become increasingly difficult. Wang said Muse Wonder can carry information across “days, weeks, months” so users don’t have to re-explain themselves. Its agent also will be able to try new tactics when it runs into problems or dead ends. The result, Wang said, is that Meta’s consumer AI agent will be able to handle more complex, even “ambiguous” tasks in addition to simple requests.

Wang and his Meta Superintelligence Labs team are already looking beyond Muse Wonder. He said MSL has begun work on “larger and more scaled models.” At the same time, the team is trying to shrink the cost of serving that technology to billions of people. Muse Wonder is larger and more expensive to operate than Meta’s smaller Muse Spark models, Wang said, so the company is already developing specialized versions that it hopes will retain much of Wonder’s capabilities while being efficient enough to deploy at massive scale.

For now, Muse Wonder is available to developers through the Meta Model API and Muse Code.

Asked how he would convince skeptics who still doubt Meta has reached the AI frontier, Wang said they should try the model and judge for themselves.

He said he’s not worried about the “haters,” which he regarded with an expletive.