Meta unveils first AI model from superintelligence team
Meta Platforms unveiled Muse Spark, the first artificial intelligence model of a costly team it established in 2025 to catch up with its competitors in the artificial intelligence race.
US tech giants are under pressure to prove that their big AI spend will pay off.
The stakes are particularly high for Meta after it hired Scale AI CEO Alex Wang in a US$14.3bn ($20.3bn) deal in 2025 and offered packages worth hundreds of millions of dollars to some engineers to build a new superintelligence team in a bid to return it to the top ranks of the AI world after a disappointing showing with its Llama 4 models in early 2025.
Superintelligence refers to artificial intelligence machines that can think better than humans. The Muse Spark is the first of this team’s new model series, known internally as the Avocado.
The company’s first model to launch in about a year will initially be available only on the little-used Meta AI app and website. The company said it will replace existing Llama models that support chatbots from WhatsApp, Instagram, Facebook and Meta’s smart glasses collection in the coming weeks.
Meta did not disclose the size of Muse Spark, a key metric often used to compare an AI system’s computing power with competitors. It has also reversed course on previous public releases of its Llama models, instead sharing only a “private preview” of Muse Spark with unnamed partners.
“This first model is small and fast by design, yet capable enough to answer complex questions in science, math, and healthcare. This is a strong foundation, and the next generation is already in development,” the company said in a blog post.
Independent evaluations of Muse Spark’s performance have shown that it catches up to the best of market leaders Google, OpenAI, and Anthropic in some areas, such as language and visual understanding, but lags behind in other areas, such as coding and abstract reasoning.
The model ranked fourth in a broad index of AI tests compiled by rating firm Artificial Analysis.
Meta CEO Mark Zuckerberg had tempered early performance expectations in January, telling investors he thought the team’s early models “will be good, but more importantly will demonstrate the rapid trajectory we’re on.”
“I expect us to push the boundaries steadily throughout the year as we continue to launch new models,” he said.
Wang, who leads the new superintelligence team, acknowledged in a series of social media posts on Wednesday that “there are definitely rough spots in the model behavior that we will iron out over time.”
He said larger versions of the model are in development and Meta plans to openly release at least some of them.
With this release, Meta has given a clearer idea of how it aims to use its models to make money and introduced shopping features built into the Meta AI chatbot that direct users directly to products they can purchase.
Overall, the company believes that applying AI to daily personal tasks will increase engagement among the more than 3.5 billion users on social media platforms, potentially giving it an advantage over rivals with smaller reach.
Muse Spark can also help users with tasks like estimating the calories in a meal from a photo or placing an image of a mug on a shelf to see what it looks like, the company said.
