Scale AI’s Muse Spark
Scale AI unveiled Muse Spark, the first model from its rebuilt MSL stack, and said the new model is now powering Meta AI after nine months of infrastructure and pipeline work. The release signals Scale's push to ship production-grade models and deepen partnerships with hyperscalers and large platform customers. (x.com)
Nine months after Meta hired Scale AI founder Alexandr Wang and put $14.3 billion into Scale for a 49% stake, the first model from Wang’s new team is already live inside Meta AI. Meta says Muse Spark now powers the Meta AI app and website, with WhatsApp, Instagram, Facebook, Messenger, and Meta’s AI glasses next in line. (scale.com) (about.fb.com) This was not a normal model launch. Meta created Meta Superintelligence Labs in 2025 after a reset of its artificial intelligence effort, and Muse Spark is the first model to come out of that rebuilt group. (techcrunch.com) (about.fb.com) Scale AI matters here because it did not start as a model company. Scale built its business on the less glamorous layer underneath artificial intelligence: turning raw text, images, and audio into training data, then grading models with human feedback and custom evaluations. (scale.com 1) (scale.com 2) That work put Scale inside the pipelines of major model builders long before it had its own headline model. In July 2024, Scale said it helped Meta’s Llama 3.1 with supervised fine-tuning, reinforcement learning from human feedback, and enterprise evaluations. (scale.com) So Muse Spark is a shift in position. The company that became known for supplying the picks and shovels is now tied to the thing at the top of the stack: a production model serving hundreds of millions of Meta users. (scale.com) (about.fb.com) Meta is pitching Muse Spark less like a research demo and more like a system built for consumer products. The company says the model is purpose-built for Meta’s apps, and that future features will cite recommendations and content people share across Instagram, Facebook, and Threads. (about.fb.com) The technical bet is speed under real traffic. Meta says Muse Spark uses multiple agents working on the same problem in parallel, which is its way of spending more test-time reasoning without making users wait as long for an answer. (techcrunch.com) That fits a bigger pattern inside Meta. In March 2026, Meta’s engineers described a new serving stack for large language model-scale systems that was built around request-centric architecture, hardware-aware design, and infrastructure tuned for sub-second latency at Meta’s size. (engineering.fb.com) Scale’s role is now easier to read. After Wang moved to Meta in June 2025, Scale said the commercial relationship with Meta would expand, but also stressed that Scale would stay independent and keep customer systems and confidential data separate. (scale.com 1) (scale.com 2) What changed this week is that the partnership is no longer just about data and evaluations behind the scenes. The first visible product from the Wang-Meta tie-up is now in front of users, and Meta is using it as proof that its rebuilt artificial intelligence organization can ship faster than the old one did. (about.fb.com) (techcrunch.com) For Scale, that turns the company from a backstage contractor into part of the story customers tell about buying artificial intelligence. For Meta, it is the first public return on a June 2025 deal that gave it a large stake in Scale and the executive who used to run it. (scale.com) (about.fb.com)
