Meta debuts first AI model from new superintelligence group

Meta Platforms Inc. debuted its latest artificial intelligence model on Wednesday — its first since Chief Executive Officer Mark Zuckerberg embarked on a multibillion-dollar overhaul of the company’s AI organisation to keep pace with rivals.

The highly anticipated model, known as Muse Spark, was created by Meta Superintelligence Labs, the new team of expensive AI researchers helmed by Chief AI Officer Alexandr Wang. Muse Spark will power the Meta AI chatbot, and in a pivot from the company’s prior open-source strategy, is a closed model, meaning its design and code won’t be made public.

Meta shares jumped 6% in New York following the announcement.

Also Read: US court declines to block Pentagon's Anthropic blacklisting for now

The model is the first major test for Zuckerberg’s new AI lab, known as MSL. The Facebook founder recruited Wang as part of a $14 billion investment into Scale AI last year after a series of setbacks left him frustrated that the company wasn’t keeping pace with competitors like OpenAI, Anthropic PBC, and Alphabet Inc.’s Google. Zuckerberg has since spent billions on AI talent and committed tens of billions more on infrastructure like data centres to catch up.

Meta has tried to keep its AI division nimble, giving researchers autonomy and minimising its typical management-heavy organisational structure, a company executive said. Wang has around 100 direct reports, said the executive, who asked not to be named, discussing internal matters.

The executive acknowledged that Muse Spark wasn’t as capable in some areas as OpenAI’s ChatGPT, Anthropic’s Claude or Google’s Gemini, but said that it was early in the company’s execution. The model is “an early data point on our trajectory,” with several larger models in development, Meta said in a blog post.

Muse Spark, known internally as Avocado during development, was built over nine months and is seen by executives as a refresh of Meta’s AI strategy, which was previously focused on its open-source Llama models. Wang is a proponent of closed models, and while Meta still plans to build open-source ones in the future, it is also considering selling API access to Muse Spark, according to the company executive.

The Meta AI chatbot will remain free for users, but the company is considering implementing subscription fees in the future, the executive said.

The name Muse Spark hints at the broader pipeline of AI offerings to come: The model is “sparking,” or jump-starting, MSL’s efforts, the executive said, describing it as small and fast. It is the first in a series of expected “Muse” models — an homage to Zuckerberg’s interest in Greek and Roman mythology.

Muse Spark was trained using several third-party open-source models, including Qwen from the Chinese tech giant Alibaba Group Holding Ltd., as well as ones from OpenAI and Google. Using a Chinese model runs counter to the stance taken by US policymakers and tech executives, who have raised concerns about China’s position in a fast-moving AI race — citing national security risks. Some of the largest American AI companies have tried to crack down on Chinese models similarly trained on US technology.

“Like others across the industry, Meta uses techniques like distillation with strict safeguards in place to learn from openly available AI models and improve our own,” said a Meta spokesperson.

Like its competitors, Muse Spark offers several levels of reasoning: “Instant” mode, “Thinking” mode and “Contemplating” mode — the last providing research-grade responses. The Meta executive said the model is good at answering questions about science, health and math, but weaker at coding.

The model will be used to improve Meta’s fleet of apps, from Instagram to Facebook to WhatsApp. It is already powering an early-stage shopping agent that aims to help users better search for products like clothing or furniture. The agent cannot yet make purchases on a user’s behalf and is still in early stages, the executive said.