Alexandr Wang Defends Meta’s Muse Spark as an ‘Appetizer’ in Bigger A.I. Push

In early April, Meta unveiled Muse Spark, its latest A.I. model and the first major release under Chief A.I. Officer Alexandr Wang. The model performs competitively on some benchmarks but still trails OpenAI’s GPT-5.4 Pro and Google’s Gemini 3.1 Pro, fueling questions about whether Meta has fallen behind. Wang believes that framing misses the point.

“The new Muse Spark model that we released is not at the tier of the leading frontier models,” Wang said during an onstage interview at Bloomberg Tech Summit in San Francisco yesterday (June 4). “But we believe it’s a very exciting data point on the trajectory, and we expect the upcoming models we release to be quite competitive with the leading models in the world.”

He described Muse Spark as an “appetizer.” Asked when the entrée would arrive, Wang replied: “We’re cooking it. We’re seeing very exciting and promising results in the process of training it right now.”

Muse Spark marks a shift for Meta. It is the company’s first proprietary model and is deployed only in Meta’s products, rather than being released openly as prior systems were. The model is designed to handle text, images, video and audio, and to support more complex, multi-step tasks, including shopping features tied to Instagram and Facebook content.

The release follows a difficult stretch for Meta’s A.I. efforts. Llama 4, launched in April 2025, was widely criticized. Two months later, Mark Zuckerberg hired Wang to lead its newly formed Superintelligence Labs and reset its strategy.

Wang said the group is focused on scaling: expanding data, computing power and research to drive improvements. And Muse Spark sits early in that process.

The barrier to the frontier isn’t money, Wang said. “It’s about continuing to scale the data, the compute … as well as continuing to scale with research. All of the labs are dramatically scaling up their models, and we’re on a much faster trajectory because we’ve been doing all this work over the past year.”

Meta is backing that approach with heavy spending. The company expects capital expenditures of $125 billion to $145 billion in 2026, up from $72.2 billion in 2025, and is targeting more than 1.3 million GPUs and roughly one gigawatt of A.I. computing capacity.

The shift to a closed model also reflects safety concerns. During development, Muse Spark triggered internal alerts, including around potential biological risks.

“When the company launches a model in a product, we have a lot of ways to mitigate some of these risks,” he said. “It’s much harder to do that when you open-source the model.”

Meta hasn’t abandoned open-source A.I. entirely and continues to develop models it considers safe to release. Whether its Llama brand will continue remains undecided. “We have exciting debates about branding internally,” Wang said, “and nothing to share right now.”

Wang said Muse Spark’s strengths are in multimodal capabilities, health-related applications and creative coding—such as generating simple games or digital tools. Those areas underpin Meta’s broader push into A.I. agents.

The company is “doubling down” on agents, aiming to build what Wang called “the best personal agents for everybody around the world.” He said he uses such tools himself to manage his health and stay in touch with friends.

The push comes alongside major internal changes. In May, Meta notified roughly 8,000 employees of layoffs and reassigned about 7,000 others to A.I.-focused roles as part of a broader reorganization.

“It’s incredibly difficult to say goodbye to teammates,” Wang said. “We don’t take any of it lightly.”