Meta's Muse Spark launch

Meta just changed course on its most famous artificial intelligence habit. Muse Spark, released on April 8, is a closed model from Meta Superintelligence Labs, not an open Llama release, and it now powers the Meta AI app and website. (about.fb.com) (bloomberg.com) A model is the engine inside a chatbot, and Meta says this one was built to handle both text and images from the start instead of bolting vision on later. Meta also says Muse Spark is “small and fast by design,” which usually means lower cost and quicker replies, even if it is not the company’s biggest system yet. (about.fb.com) (cnbc.com) The timing tells you what this launch is really about. Meta says Superintelligence Labs rebuilt its artificial intelligence stack over the last nine months, and Muse Spark is the first public product from that reset. (about.fb.com) (cnbc.com) That reset started after Mark Zuckerberg brought in Alexandr Wang through Meta’s roughly $14.3 billion Scale AI deal in June 2025. Bloomberg and CNBC both describe Muse Spark as the first major test of the new lab Wang now leads. (cnbc.com) (bloomberg.com) Meta’s pitch is not “we built the biggest brain.” Meta’s pitch is that Muse Spark can reason through harder questions, can look at photos as well as text, and can split a task across multiple subagents, like one helper planning flights while another compares destinations. (about.fb.com) That product angle fits Meta better than a bare benchmark race. The company says Muse Spark will roll out beyond the Meta AI app and website to WhatsApp, Instagram, Facebook, Messenger, and its artificial intelligence glasses in the coming weeks. (about.fb.com) Meta is also opening a second business door with this model. The company says it will offer private application programming interface preview access to select partners, and Bloomberg reports Meta is considering selling access while keeping the consumer chatbot free. (about.fb.com) (bloomberg.com) The awkward part is that Muse Spark does not appear to win everywhere. Bloomberg reported that a Meta executive acknowledged the model was not as capable as OpenAI’s ChatGPT, Anthropic’s Claude, or Google’s Gemini in some areas, while outside coverage says it looks stronger on language than on coding. (bloomberg.com) (theguardian.com) That helps explain Meta’s hiring map. If language and consumer chat are getting better but coding still lags, the company has a reason to keep splitting talent across reasoning research, coding-agent work, and safety instead of betting everything on one giant general model. (theguardian.com) (about.fb.com) The bigger shift is strategic, not cosmetic. Meta spent years making Llama the symbol of its open approach, and Muse Spark is the clearest sign yet that Zuckerberg now wants a mixed playbook: open models where that helps distribution, closed models where that helps speed, control, and product integration. (bloomberg.com) (cnbc.com) For now, Muse Spark looks less like a final answer than a checkpoint. Meta calls it an “early data point,” says larger Muse models are already in development, and is using this smaller release to prove the rebuilt lab can ship something real after a very expensive year of recruiting and reorganization. (about.fb.com) (bloomberg.com)