Meta debuts new AI model in first test of costly ‘superintelligence’ team

Meta on Wednesday unveiled Muse Spark, the first artificial intelligence model from a costly team it assembled last year to catch up with rivals in the AI race.

US tech companies are under pressure to prove their huge AI outlays will pay off. The stakes are especially high for Meta after it hired Alex Wang, Scale AI CEO, last year in a $14.3bn deal and offered some engineers pay packages of hundreds of millions of dollars to staff a new “superintelligence” team, an attempt to propel itself back into the AI world’s top ranks after a disappointing showing with its Llama 4 models early last year. Superintelligence refers to AI machines that could outthink humans. Muse Spark is the first in a new series of models, known internally as Avocado, from that team.

The model, the first the company has released in about a year, initially will be available only on the lightly used Meta AI app and website. In the coming weeks, it will replace the existing Llama models powering chatbots on WhatsApp, Instagram, Facebook and Meta’s collection of smart glasses, the company said.

Meta did not disclose Muse Spark’s size, a key measure typically used to compare an AI system’s computing power with rivals. It also changed course from previous open releases of its Llama models, instead sharing only a “private preview” of Muse Spark with unnamed partners.

“This initial model is small and fast by design, yet capable enough to reason through complex questions in science, math and health. It is a powerful foundation, and the next generation is already in development,” the company said in a blogpost.

Independent evaluations of Muse Spark’s performance showed it catching up with top models from market leaders Google, OpenAI and Anthropic in some areas, such as language and visual understanding, but lagging in others, like coding and abstract reasoning.

The model tied for fourth place on a broad index of AI tests compiled by the evaluation firm Artificial Analysis.

Mark Zuckerberg, Meta’s CEO, had tempered expectations for early performance, telling investors in January that he thought the team’s first models “will be good but, more importantly, will show the rapid trajectory that we’re on”.

“I expect us to steadily push the frontier over the course of the year as we continue to release new models,” he had said.

Wang, who runs the superintelligence team, acknowledged in a series of social media posts on Wednesday that “there are certainly rough edges we will polish over time in model behavior.” He said bigger versions of the model were in development and that Meta was planning to release at least some of them openly.

With the release, Meta gave a clearer sense of how it aims to use its models to make money, teasing shopping features embedded within its Meta AI chatbot that point users directly to products they can buy.

Broadly, the company is betting that applying AI to everyday personal tasks will boost engagement among the more than 3.5 billion users across its social media platforms, potentially giving it an edge over rivals with a smaller reach.

Muse Spark can also help users with tasks such as estimating the calories in a meal from a photo or superimposing an image of a mug on a shelf to see how it looks, the company said.

An extra Contemplating Mode, which runs multiple agents simultaneously to boost reasoning power, would allow Muse Spark to take on the extended thinking modes of Google’s Gemini Deep Think and OpenAI’s GPT Pro.

Meta said people could use the mode for efficiently planning a family vacation, having one agent draft a travel itinerary while the other looks up kid-friendly activities.