Meta’s Muse Spark arrives
Meta revealed Muse Spark, a new proprietary AI model that looks like a strategic shift away from relying solely on the open-weight Llama family. Coverage says Muse Spark is being positioned for efficiency and performance gains under the Superintelligence Labs initiative, marking Meta’s first major model since that leadership change ( ).
Meta just put its newest artificial intelligence model inside its own apps first, and that alone tells you what changed. 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 scheduled to get it in the following weeks. (about.fb.com) For three years, Meta’s artificial intelligence story was mostly Llama, a family of models whose weights were released for others to download and build on. Muse Spark is different: Artificial Analysis says it is Meta’s first frontier model that is not being released as open weights. (artificialanalysis.ai, techcrunch.com) That makes Muse Spark less like a public blueprint and more like a private engine bolted directly into Meta’s products. Meta said it will offer private preview application programming interface access to select partners, but the first destination is Meta’s own assistant, not the broader developer world. (about.fb.com, cnbc.com) The background is a rough stretch for Meta’s last big model cycle. CNBC reported that Muse Spark is Meta’s first major model since the company brought in Scale AI founder Alexandr Wang in June 2025, after Llama 4 failed to create the kind of developer excitement Meta wanted. (cnbc.com, techcrunch.com) Meta then reorganized its artificial intelligence work under Meta Superintelligence Labs, the unit Wang now leads. In its launch post, Meta said that team spent the last nine months rebuilding its artificial intelligence stack “from the ground up,” and called Muse Spark the first model in a new Muse series. (about.fb.com, techcrunch.com) Meta is not selling Muse Spark as the biggest model on earth. It is selling it as a smaller, faster model that can still reason through harder questions, which is the difference between a race car and a delivery van: one wins headlines, the other has to show up all day inside products people actually use. (about.fb.com, cnbc.com) That efficiency claim is the part rivals will watch closely. VentureBeat reported that Meta says Muse Spark reaches its reasoning performance with more than an order of magnitude less compute than Llama 4 Maverick, which was Meta’s previous mid-size flagship. (venturebeat.com, huggingface.co) Outside testing suggests this is not just a packaging change. Artificial Analysis said Muse Spark scored 52 on its Intelligence Index, far above the 18 scored by Llama 4 Maverick at release, and placed within the top five models it has benchmarked. (artificialanalysis.ai) Meta is also using Muse Spark to change what its assistant does on screen. The company said Meta AI can now switch modes for quick answers or harder reasoning, launch multiple subagents in parallel for tasks like trip planning, and use image understanding so a person can point a camera at a shelf or product and ask questions about what they see. (about.fb.com) So the real shift is not just a new model name. Meta spent 2023 through 2025 teaching the market to think “Llama” when it talked about artificial intelligence, and on April 8, 2026, it started teaching a different lesson: the models that matter most to Meta may now be the ones it keeps closest. (techcrunch.com, about.fb.com, artificialanalysis.ai)
