Meta debuts Muse Spark
Meta unveiled a new, proprietary multimodal AI called Muse Spark that the company is positioning as a serious rival to Google and OpenAI. The model emphasizes multimodal reasoning and a “Contemplating mode,” and early coverage highlights strong benchmark performance and a shift away from Meta’s prior open-model posture. That signals hiring teams at big tech will likely probe how candidates evaluate models, benchmark results, and product trade-offs rather than just citing leaderboard numbers. (cnbc.com) (wired.com)
Meta just put its newest artificial intelligence model inside the Meta AI app and website first, instead of releasing it the way it released Llama. On April 8, 2026, the company said Muse Spark is already live in Meta AI and is coming to WhatsApp, Instagram, Facebook, Messenger, and its artificial intelligence glasses in the next few weeks. (about.fb.com) That is a sharp turn for a company that spent the last two years telling developers its open Llama models were the safer bet against closed systems from OpenAI and Google. Muse Spark will get only a private application programming interface preview for selected partners, which means most outsiders cannot inspect or build on it the way they could with Llama weights. (about.fb.com) (cnbc.com) Meta says the model was built over the last nine months by Meta Superintelligence Labs, the new group led by Alexandr Wang after Meta’s $14.3 billion investment in Scale AI in June 2025. CNBC reports Muse Spark was originally codenamed Avocado and is the first major model from that team. (cnbc.com) The company is not pitching Muse Spark as the biggest model on earth. It is pitching it as small and fast, with enough reasoning power to answer harder questions in science, math, and health without the long lag that turns a chatbot into a loading screen. (about.fb.com) (cnbc.com) Meta’s key feature is something it calls Contemplating mode. In plain English, that means the assistant can split one job into smaller jobs, send them to multiple internal helpers at once, and merge the answers, like handing a vacation plan to three coworkers instead of one. (about.fb.com) (dnyuz.com) The second bet is vision. Meta says Muse Spark is built to look at photos and products directly, so a person can point a camera at an airport snack shelf, ask for the highest-protein option, and get an answer without typing every label into a text box. (about.fb.com) That matters most in Meta’s own hardware and apps, because cameras, shopping posts, and friend recommendations are already there. Meta says Muse Spark will eventually cite recommendations and content shared across Instagram, Facebook, and Threads, which gives it a home-field advantage that outside chatbot makers do not have. (about.fb.com) The launch also lands after a stumble. CNBC says Meta was trying to recover from the weak reception to its Llama 4 family, which did not win over developers the way earlier Llama releases did, and Zuckerberg responded by rebuilding the stack and changing course. (cnbc.com) So the real story is not only that Meta has a new model. It is that Meta is acting more like the rivals it used to criticize: keeping the strongest system close, wiring it into products with billions of users, and offering outside access slowly and on Meta’s terms. (about.fb.com) (cnbc.com)
