Scale AI’s Muse Spark to power Meta

# Scale AI’s Muse Spark to power Meta Scale AI says the first model from its rebuilt artificial intelligence stack is already running inside Meta AI after nine months of work. The claim came with the debut of Muse Spark, which Scale described as the first model produced from a new foundation of infrastructure, model architecture, and data pipelines. (cnbc.com) That is a sharp turn for a company best known for supplying the parts behind other people’s artificial intelligence systems. Scale built its business on data labeling, evaluation, fine-tuning, and safety work for model makers, not on shipping a consumer-facing model of its own. (scale.com 1) (scale.com 2) Muse Spark matters because it suggests Scale wants to move one layer up the artificial intelligence stack. Instead of only helping labs prepare training data and test outputs, Scale is now presenting itself as a company that can build a deployable model that plugs directly into a platform with billions of users. (scale.com) (cnbc.com) The timing also connects to the deepening relationship between Scale and Meta. On June 12, 2025, Scale announced a major Meta investment that valued Scale at more than $29 billion, expanded the commercial relationship between the two companies, and sent founder Alexandr Wang to Meta to work on its artificial intelligence efforts. (scale.com) Meta’s side of the story is that Muse Spark is the first major model from the team Wang now leads. CNBC reported on April 8, 2026 that Meta described Muse Spark as the first model in a new Muse series developed by Meta Superintelligence Labs, the group overseen by Wang, and said the model was built over the past nine months. (cnbc.com) (axios.com) That creates an unusual picture around ownership and positioning. Public reporting describes Muse Spark as a Meta model led by Wang’s team inside Meta, while the story framing from Scale emphasizes that the model came from a rebuilt stack associated with Scale and is now powering Meta AI. Based on the available public material, the most precise reading is that the companies are presenting the same launch from two angles: Meta as the platform owner and product shipper, Scale as the upstream builder and infrastructure partner. (cnbc.com) (scale.com) That distinction matters in today’s artificial intelligence market because the most valuable layer is no longer just the chatbot that consumers see. There is also a growing business in the hidden layer underneath: the data systems, evaluation loops, safety controls, and model optimization work that determine whether a model can actually be trusted at scale. Scale has spent years building its brand in that hidden layer. (scale.com 1) (scale.com 2) Meta, meanwhile, has been looking for a faster path back into the front rank of model makers. CNBC reported that Muse Spark follows a period in which Meta was trying to regain momentum after the underwhelming reception to its previous Llama 4 family, and that Meta is highlighting Muse Spark as small, fast, and competitive rather than as a giant frontier model. (cnbc.com) That “small and fast” positioning is important. Models embedded into products like Meta AI, messaging tools, or wearable devices often need low latency and lower serving costs more than they need to win every benchmark, because users notice delay immediately and platforms pay for every inference. Meta’s own product push around Meta AI and artificial intelligence glasses shows why efficient models are strategically useful. (cnbc.com) (meta.com) Scale has also gone out of its way to say that the Meta relationship does not make it a Meta captive. In June 2025, Scale said it would remain an independent company, would not integrate operations with Meta, and would not give Meta access to Scale internal systems or customer confidential information. (scale.com) That independence claim is central to how Scale wants this launch understood. If Scale can help build a model that lands inside Meta AI while still serving governments, enterprises, and other model labs, it starts to look less like a contractor and more like a neutral supplier of artificial intelligence infrastructure. (scale.com 1) (scale.com 2) The bigger takeaway is that Scale appears to be testing a new identity in public. For years, the company sold picks and shovels to the artificial intelligence gold rush; with Muse Spark, it is signaling that the picks and shovels business can also produce the engine itself. (scale.com) (cnbc.com) There is still a lot that has not been disclosed. Neither the public reporting nor Scale’s existing public materials, from what is available so far, spell out Muse Spark’s full architecture, parameter count, training mix, benchmark table, or the exact scope of where it is already powering Meta AI. (cnbc.com) (axios.com) What is clear on April 8, 2026 is the direction of travel. Meta is using Muse Spark to show it can still ship new models quickly under Wang, and Scale is using the same moment to show it can be more than a data-and-evaluation company. (cnbc.com) (scale.com)