Muse Spark powering Meta

Meta just swapped the engine inside Meta AI. On April 8, 2026, Meta said a new model called Muse Spark now powers the Meta AI app and the meta.ai website, replacing the earlier setup built around Llama 4 in those products. (about.fb.com) That is a fast change for a product used across Meta’s consumer software. When Meta launched the Meta AI app on April 29, 2025, the company said the app was built with Llama 4, which made Llama the public face of Meta AI at the time. (about.fb.com) Muse Spark is not a chatbot app by itself. Meta describes it as the first model in a new Muse series, which means it is the core prediction system underneath the assistant, like swapping the motor inside a car while keeping the dashboard people already know. (about.fb.com) A model is the part that turns a question into an answer. If you ask for a trip plan, the model is the thing deciding which facts to pull together, which steps to take, and which words to send back. (about.fb.com) Big consumer assistants live or die on speed. Meta said Muse Spark was designed to be “small and fast,” which usually means a company is trying to cut the delay between your prompt and the first useful answer without giving up too much reasoning ability. (about.fb.com) The hard part is that speed alone is not enough. Meta also said Muse Spark can handle complex questions in science, math, and health, so the company is claiming it found a balance between quick replies and more demanding reasoning tasks. (about.fb.com) Another piece of the puzzle is multimodal perception. That phrase means the system can work with more than typed text, so a person can point a camera at a shelf or a product and ask the assistant to interpret what it sees. (about.fb.com) Meta tied Muse Spark to a more agent-like version of Meta AI. The company said the assistant can launch multiple subagents in parallel, with one drafting an itinerary while another compares destinations and a third searches for activities, which is Meta’s way of turning one question into several smaller jobs at once. (about.fb.com) Now the company story gets more interesting. Meta said Meta Superintelligence Labs rebuilt its artificial intelligence stack from the ground up over the last nine months, and Muse Spark is the first model to come out of that rebuilt system. (about.fb.com) An artificial intelligence stack is the full factory, not just the model. It includes the training software, the chips and servers, the testing systems, the safety checks, and the data pipelines that move raw information into training-ready form. (about.fb.com) (scale.com) Data pipelines are the conveyor belts in that factory. They collect data, clean it, label it, score it, and route it into training and evaluation systems, which is why companies like Scale AI built businesses around supplying data, evaluations, and human feedback to model makers. (scale.com 1) (scale.com 2) That is where the Scale AI angle comes in. Scale has spent years selling data and evaluation services to major model builders, and the company said in 2024 that it supplied data to organizations including Meta, OpenAI, and Microsoft. (scale.com) The relationship between Meta and Scale got much closer in June 2025. Scale said Meta made an investment that valued Scale at more than $29 billion, expanded the companies’ commercial relationship, and brought Scale founder Alexandr Wang over to work on Meta’s artificial intelligence efforts. (scale.com) Scale also spent the weeks after that deal telling customers it was still separate. On June 18, 2025, Scale said it would remain independent, would not give Meta access to Scale internal systems, and would keep customer confidentiality protections in place. (scale.com) So when Wang described Muse Spark as a milestone for rebuilt infrastructure and data pipelines, he was pointing at the plumbing behind the model, not just the model itself. The notable part is that Meta says this first Muse model is already in production on the Meta AI app and website, which suggests the new stack moved from rebuild to live consumer deployment unusually quickly. (about.fb.com) Meta is also not treating this as a one-off release. The company said larger Muse models are already in development, and it plans to roll Muse Spark out beyond the app and website to WhatsApp, Instagram, Facebook, Messenger, and artificial intelligence glasses in the coming weeks. (about.fb.com) That makes Muse Spark less like a finished flagship and more like the first working car off a retooled assembly line. The headline is not only that Meta has a new model on April 8, 2026, but that its rebuilt stack produced something small enough to ship fast, strong enough to replace a live assistant, and central enough to become the base layer for Meta AI across several products. (about.fb.com)