Scale’s Muse Spark Debuts

Nine months after rebuilding its artificial intelligence stack from the ground up, Meta put a new model called Muse Spark into the Meta AI app and the meta.ai website on April 8, 2026, and said it will roll out to WhatsApp, Instagram, Facebook, Messenger, and its artificial intelligence glasses in the coming weeks. (about.fb.com) That launch is tied directly to Scale AI founder Alexandr Wang, who joined Meta’s artificial intelligence effort in June 2025 after Meta invested $14.3 billion for a 49% stake in Scale. (scale.com, techcrunch.com) Scale used to be best known as the company behind the plumbing of artificial intelligence: data labeling, evaluations, reinforcement learning from human feedback, and enterprise deployment tools. In July 2024, Scale and Meta were still publicly talking about helping companies customize and test Meta’s Llama 3.1 model for business use. (scale.com, scale.com) Now the relationship looks different. Meta says Muse Spark is the first model from Meta Superintelligence Labs, the unit Wang leads, and says the model is already powering Meta AI for consumers. (about.fb.com, techcrunch.com) Meta describes Muse Spark as “small and fast by design,” which is the artificial intelligence version of building a compact engine that can start instantly instead of a bigger engine that takes longer to spin up. Meta says that smaller model still handles science, math, health, and image-heavy tasks, and it plans to offer private application programming interface access to select partners. (about.fb.com) The trick is not just the model. Meta says Meta AI can now launch multiple subagents in parallel, meaning several smaller workers tackle one question at the same time, like splitting a research project across a team instead of handing it to one person. (about.fb.com, techcrunch.com) That helps explain why this story reaches beyond one product launch. Big customers do not just buy a model anymore; they buy the full stack of data pipelines, testing systems, security controls, application tools, and inference infrastructure that keeps the model reliable after launch. (scale.com, scale.com) Scale has been selling exactly that enterprise layer for years, including private evaluations, model monitoring, and deployment tools for legal, finance, and education use cases. The new wrinkle is that the company most associated with those support layers now has its founder helping run a top-tier model program inside one of Scale’s biggest partners. (scale.com, scale.com) Meta and Scale both spent June 2025 insisting that Scale would remain independent, that Meta would not get access to Scale’s internal systems, and that customer information would stay isolated. Those statements matter more now that Meta has shipped a model from the team run by Scale’s founder and tied it to a product used across several Meta apps. (scale.com, scale.com) Muse Spark also marks a shift in Meta’s posture. In 2024, Meta was still emphasizing open research and openly released models like Llama, but Muse Spark is being introduced as a product-first model built for Meta’s own assistants, interfaces, and devices. (about.fb.com, about.fb.com) So the real news is not only that a new model appeared on April 8. It is that the old line between “the company that builds the model” and “the company that builds the infrastructure around the model” is getting blurrier, and Meta is betting that the team around Alexandr Wang can do both at consumer scale. (about.fb.com, scale.com)