Meta unveils Muse Spark
Meta says it has a new multimodal model called Muse Spark that will power capabilities across its apps, from shopping recommendations to health-image queries. That suggests consumer discovery and platform surfacing will increasingly be mediated by in-house multimodal AI rather than just third‑party APIs. (thekeyword.co)
Meta is no longer treating artificial intelligence like a side feature inside its apps. On April 8, 2026, it said a new in-house model called Muse Spark already runs the Meta AI app and website, with WhatsApp, Instagram, Facebook, Messenger, and its artificial-intelligence glasses next in line. (about.fb.com) A model is the engine underneath a chatbot, image tool, or recommendation box. The company says Muse Spark is “small and fast by design,” which means it is built to answer quickly enough for consumer apps people open dozens of times a day. (about.fb.com) Meta is calling this a multimodal system, which means one model can handle different kinds of input instead of just text. In practice, the company says that includes tasks like shopping help and answering questions about images, including health-related images. (about.fb.com) That is a shift from the earlier phase of the artificial-intelligence boom, when many consumer apps plugged into outside models through application programming interfaces, which are rented software connections. Meta says Muse Spark will also be offered in private preview through an application programming interface, but the first priority is clearly Meta’s own products. (about.fb.com) The company is tying the model directly to discovery inside its social apps. Meta says Muse Spark will unlock features that can cite recommendations and content people share across Instagram, Facebook, and Threads, which turns posts from friends and creators into raw material for artificial-intelligence answers. (about.fb.com) Meta has been moving toward this for a while. In September 2024, it said Meta AI had more than 400 million monthly users and added multimodal features through Llama 3.2 across Messenger, Facebook, WhatsApp, and Instagram direct messages. (about.fb.com) Then in April 2025, Meta launched a standalone Meta AI app and said it was built with Llama 4, linking the assistant across the app, the web, and its glasses companion experience. Muse Spark now replaces that earlier foundation at the app and website level, which shows how fast Meta is swapping model layers under products that already have users. (about.fb.com, about.fb.com) Meta is also using this launch to advertise a new internal lab structure. It says Muse Spark is the first model from Meta Superintelligence Labs, a group that rebuilt the company’s artificial-intelligence stack over the last nine months and plans larger Muse models after this one. (about.fb.com, about.fb.com) The practical result is that more of what people see, ask, and buy inside Meta’s apps may be filtered through Meta’s own model before they ever leave the platform. If shopping suggestions, image answers, and social recommendations all run through one in-house system, Meta gets to decide not just the feed, but the assistant sitting on top of the feed. (about.fb.com)
