Meta launches Muse Spark
Meta introduced Muse Spark, its first major model from the Superintelligence Labs, signalling a push toward higher-performance proprietary models while keeping Llama-era open deployments alive. (cnbc.com) The move looks like a dual strategy — preserve the open-model ecosystem for self-hosting customers while developing proprietary offerings for higher-tier enterprise use. (techcrunch.com)
Meta just put a new model called Muse Spark into the Meta AI app and website, and it is the first major model to come out of Meta Superintelligence Labs since that group was formed. Meta said WhatsApp, Instagram, Facebook, Messenger, and its artificial intelligence glasses will get it in the coming weeks. (about.fb.com) That is a break from the last few years, when Meta’s public artificial intelligence identity was mostly tied to Llama, the open model family that outside developers could download and run themselves. Muse Spark is instead being positioned first as a model built for Meta’s own products, with only a private application programming interface preview for selected partners. (about.fb.com, ai.meta.com) Meta says Muse Spark is its “most powerful model yet,” but the more revealing line is where it plans to use that power: recommendations and shared content pulled across Instagram, Facebook, and Threads. That points to a model designed less like a general public utility and more like an engine wired straight into Meta’s own apps and data. (about.fb.com) The corporate backdrop is a fast rebuild of Meta’s artificial intelligence operation. TechCrunch reported on April 8 that Meta recruited former Scale AI chief executive Alexandr Wang to lead Meta Superintelligence Labs and put $14.3 billion into Scale AI for a 49 percent stake. (techcrunch.com) Scale AI matters here because it sits in the unglamorous part of artificial intelligence: labeling and organizing the data that models learn from. Buying nearly half of that pipeline while hiring its founder is a way to tighten control over both the training inputs and the product roadmap at the same time. (techcrunch.com) Meta is not saying Llama is over. The company still maintains Llama documentation and responsible-use material for developers, which means the open-model track is still alive even as Muse opens a second track aimed at Meta-controlled deployment. (ai.meta.com, about.fb.com) That gives Meta two different products for two different buyers. Llama still fits companies that want to self-host models inside their own walls, while Muse Spark fits customers who want higher-end features delivered through Meta’s software and, eventually, its paid interfaces. (about.fb.com, techcrunch.com) There is also a speed story inside the launch. 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 suggests Meta is trying to make stronger models that are cheaper to run across billions of consumer interactions. (venturebeat.com) Meta previewed that direction in a January 2026 post saying the year would bring more personalized artificial intelligence products and new ways to transform how people use its services. Muse Spark looks like the first concrete product in that plan: less a downloadable model for the wider ecosystem, more a house model for Meta’s own empire. (about.fb.com, about.fb.com)
