In early April, Meta unveiled Muse Spark, its latest AI model and the first major release under AI CEO Alexander Wang. The model performs competitively on some benchmarks but still lags behind OpenAI’s GPT-5.4 Pro and Google’s Gemini 3.1 Pro, raising questions about whether Meta has fallen behind. Wang thinks the framing misses the point.
“The new Muse Spark model we launched is not at the level of the leading frontier models,” Wang said during an on-stage interview at the Bloomberg Technology Summit in San Francisco yesterday (June 4). “But we think it’s a very exciting data point down the road, and we expect the next models we launch will be fully competitive with the world’s leading models.”
Moses described Spark as an “appetiser”. When asked when the main course would arrive, Wang replied: “We are cooking it. We are seeing very exciting and promising results in his training process now.”
Muse Spark represents a shift in the Meta. It is the company’s first private model and is only published in Meta products, rather than publicly released as previous systems were. The model is designed to handle text, images, video, and audio, and support more complex, multi-step tasks, including shopping features associated with Instagram and Facebook content.
The release comes after a difficult period for Meta’s AI efforts. Llama 4, which was launched in April 2025, was widely criticized. Two months later, Mark Zuckerberg appointed Wang to lead the 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. Muse Spark sits early in the process.
Wang said the barrier to the border is not money. “It’s about continuing to scale up data and computing…as well as continuing to scale up research. All the labs are dramatically scaling up their models, and we’re on a much faster trajectory because we’ve done all this work over the past year.”
Meta supports this approach with significant 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 graphics processing units and nearly one gigawatt of AI computing capacity.
The shift to the closed model also reflects safety concerns. During development, Muse Spark raised internal alerts, including about potential biological risks.
“When a company launches a model into a product, we have a lot of ways to mitigate some of those risks,” he said. “It’s very difficult to do that when the model is open source.”
Meta has not abandoned open source AI completely, and continues to develop models that it considers safe to release. Whether or not its Llama brand will continue is yet to be decided. “We have exciting discussions about the brand internally, and we have nothing to share now,” Wang said.
Wang said Muse Spark’s strengths lie in multimedia capabilities, health-related applications and creative programming, such as creating simple games or digital tools. These areas support Meta’s broader orientation toward AI agents.
The company is “doubling down” on agents, with the goal of building what Wang called “the best personal agents for everyone around the world.” He said he uses these tools himself to manage his health and stay in touch with friends.
This push comes along with major internal changes. In May, Meta notified nearly 8,000 employees of layoffs and reassigned about 7,000 more to AI-focused roles as part of a broader reorganization.
“It is very difficult to say goodbye to our teammates,” Wang said. “We don’t take any of it lightly.”
