Meta Launches Muse Spark

Meta just put out its first big language model since April 2025, and the bigger surprise is the name: not Llama, but Muse Spark. Meta says Spark is the first model from Meta Superintelligence Labs, the new unit it built after reshuffling its artificial intelligence effort last year. (about.fb.com) This one is already live inside Meta AI on the web and in the Meta AI app, and Meta says it will roll out to WhatsApp, Instagram, Facebook, Messenger, and its artificial intelligence glasses in the coming weeks. Meta is not leading with “biggest model” claims here; it is leading with “small and fast by design.” (about.fb.com) A large language model is the prediction engine behind a chatbot. You type a question, the model guesses the next useful word over and over, the way your phone guesses the next word in a text message, except at much larger scale. (cnbc.com) Meta says Muse Spark can reason through harder questions in science, math, and health, and it describes the model as multimodal, which means it can work with images as well as text. Think of that as the difference between a clerk who only reads forms and one who can also look at the photo attached to the form. (about.fb.com) The backdrop is that Meta’s last flagship family, Llama 4, arrived on April 5, 2025, and did not reset the race the way Meta wanted. CNBC says the weak reception pushed Mark Zuckerberg to change strategy and rebuild the stack under a new team. (techcrunch.com, cnbc.com) That new team is led by Alexandr Wang, the former Scale AI chief executive who joined Meta in June 2025 as its first chief artificial intelligence officer. Meta also put $14.3 billion into Scale AI in that same push, which shows this was not a routine product update but a very expensive reset. (meta.com, cnbc.com) The other break with the past is distribution. Artificial Analysis says Muse Spark is Meta’s first frontier model that is not being released as open weights, which means outside developers do not get the downloadable model files the way they did with earlier Llama releases. (artificialanalysis.ai) On one independent benchmark from Artificial Analysis, Muse Spark scored 52 on its Intelligence Index, behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6, but far ahead of where Llama 4 Maverick landed at release. The same testing group said Muse Spark was especially strong on vision tasks and reasoning, while its agent-style work performance was less impressive. (artificialanalysis.ai) Meta says an application programming interface is coming, which would let outside developers pay to plug Muse Spark into their own apps. That would move Meta closer to the business model used by OpenAI, Anthropic, and Google, instead of treating its model mostly as a free engine for its own products. (cnbc.com, about.fb.com) So the launch is really two announcements in one. Meta shipped a faster chatbot brain for billions of people, and it quietly signaled that the Llama era of “open weights first” is no longer the whole plan. (about.fb.com, artificialanalysis.ai)