Meta unveils Muse Spark
Meta rolled out Muse Spark, its first major model from the reorganised Superintelligence Labs, touting much higher reasoning efficiency than the troubled Llama 4 lineage. Coverage says Muse Spark achieves similar reasoning with far less compute via a ‘thought compression’ approach, though Meta admits gaps remain in long-horizon agents and coding workflows (techcrunch.com (venturebeat.com).
Meta spent most of 2025 getting mocked for Llama 4, the artificial intelligence model family that was supposed to keep pace with OpenAI and Anthropic but instead landed with weak reviews and internal frustration. On April 8, 2026, it answered with a new name, a new lab, and a new model called Muse Spark. (about.fb.com) A reasoning model is a system that tries to solve problems step by step instead of blurting out the first likely sentence. The expensive part is that every extra step burns more computing power, which is why companies keep building larger chips-and-data-center budgets just to make models think longer. (venturebeat.com) Meta says Muse Spark cuts that cost with what it calls thought compression, which works like turning a long page of scratch notes into a tight outline before moving to the next step. The company says Spark reaches similar reasoning quality with more than an order of magnitude less compute than Llama 4 Maverick, its previous mid-size flagship. (venturebeat.com) The company did not just swap models; it rebuilt the organization that makes them. Muse Spark is the first major release from Meta Superintelligence Labs, the unit Mark Zuckerberg created after reshuffling Meta’s artificial intelligence effort and putting Alexandr Wang in charge as chief artificial intelligence officer. (techcrunch.com) Meta is also changing where this model lives. Muse Spark already powers the Meta AI app and website, and Meta says it will roll out to WhatsApp, Instagram, Facebook, Messenger, and its artificial intelligence glasses in the coming weeks. (about.fb.com) That product focus is a break from the Llama era, when Meta’s big pitch was open-weight models that outside developers could download and build on. This time, Meta says it will offer Muse Spark only in a private preview application programming interface to select partners while keeping the main deployment inside its own apps. (about.fb.com) Meta says Spark is natively multimodal, which means it was built to handle text, images, tools, and agent-style workflows in one system instead of bolting those abilities on later. The company is also introducing a Contemplating mode that runs multiple agents in parallel, like sending several researchers to work the same case at once and then comparing notes. (about.fb.com) The catch is that Meta is not claiming victory everywhere. In its own launch materials and interviews, the company says gaps remain in long-horizon agents, which are systems that have to remember and execute many steps over time, and in coding workflows, where rivals still have an edge. (techcrunch.com) So the real news is not that Meta has suddenly leapt past everyone else. It is that after a year of betting billions on a reorganization, Meta now has a model it says is competitive again, cheaper to run per reasoning task, and built first for the 3 billion-plus people who already use its apps rather than for the open model crowd that grew around Llama. (cnbc.com)
