Meta debuts Muse Spark

Meta just put a new artificial intelligence model called Muse Spark into its chatbot, nine months after spending $14.3 billion on Scale AI and bringing in Scale founder Alexandr Wang to run its new Meta Superintelligence Labs. Muse Spark went live on April 8 in the Meta AI app and on meta.ai, with WhatsApp, Instagram, Facebook, Messenger, and Meta’s artificial intelligence glasses next in line. (about.fb.com) A large language model is the prediction engine under a chatbot: you type a question, and the model guesses the next useful word over and over until it produces an answer. Meta says Muse Spark is “small and fast by design,” which means it is aiming for quick responses inside consumer apps, not just bragging rights on giant benchmark charts. (about.fb.com) Meta is not treating this as another Llama release. The company says Muse Spark is the first model in a new Muse family built by Meta Superintelligence Labs after what it called a nine-month rebuild of its artificial intelligence stack “from the ground up.” (about.fb.com) That reset came after Meta’s previous flagship, Llama 4, failed to win over developers the way OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude have. CNBC reported that Zuckerberg was frustrated by the weak reception and changed course, making Wang the executive in charge of the new effort. (cnbc.com) Wang arrived through an unusual deal. In June 2025, Meta bought a 49% stake in Scale AI for $14.3 billion, and Wang left the company he founded to join Meta’s “superintelligence” push while staying on Scale’s board. (cnbc.com) Muse Spark is built to work on more than text. Meta says the model can look at photos, compare products, answer visual science and math questions, and eventually do more through its glasses, which turns the assistant from a text box into something closer to a camera that can explain what it sees. (about.fb.com) Meta also says the assistant can launch multiple subagents in parallel, which is like sending several interns to research the same problem at once and then combining their notes. TechCrunch reported that Meta plans a “Contemplating” mode for harder questions, using that parallel setup to spend more time reasoning without making replies feel too slow. (techcrunch.com) The distribution plan is the real weapon here. OpenAI had to build ChatGPT into a habit from scratch, while Meta can drop Muse Spark into apps used by billions of people and then offer a private application programming interface preview to selected partners later. (about.fb.com) Meta is also steering away from a pure open-source story. CNBC reported that the company wants to sell access to Muse Spark through an application programming interface, which would give developers a way to build on Meta’s model and give Meta a new revenue stream beyond advertising. (cnbc.com) There is a catch in the product design: TechCrunch reported that users need a Meta account such as Facebook or Instagram to log in, and Meta is pitching the system as a “personal superintelligence” assistant. That combination raises the obvious question of how much of your existing Meta identity could shape the answers, recommendations, and memory the assistant gives back. (techcrunch.com) So this launch is less “Meta built the smartest model” than “Meta rebuilt its artificial intelligence team, hired a new boss through a $14.3 billion deal, and now has a new engine ready to spread across Facebook, Instagram, WhatsApp, Messenger, the web, and glasses.” If Muse Spark works well enough in all those places at once, Meta does not need to beat every rival in the lab to become hard to avoid in daily life. (cnbc.com)