Meta’s Muse Spark debut

Meta just put a new brain inside its apps, and it is not the Llama system the company spent years telling developers to build around. On April 8, Meta said Muse Spark already powers the Meta AI app and website and is rolling out to WhatsApp, Instagram, Facebook, Messenger, and its artificial intelligence glasses in the coming weeks. (about.fb.com) That is a sharp change in where Meta’s artificial intelligence work shows up first. Instead of starting with a broad developer release, Meta said Muse Spark is “purpose-built for Meta’s products” and only a private application programming interface preview will go to select partners. (about.fb.com) For the last few years, Meta’s main pitch was Llama, a family of model weights it released for outside developers to download, fine-tune, and run on their own systems. Meta’s own responsible use materials said the company was committed to open science and a “vibrant AI-innovation ecosystem,” which made openness part of the brand. (ai.meta.com, ai.meta.com) Meta reinforced that message as recently as LlamaCon in April 2025. At that event, the company introduced a Meta AI app, previewed a Llama application programming interface, and framed Llama as the center of a growing outside ecosystem. (forbes.com) Muse Spark points the other way. Bloomberg reported that Muse Spark is a closed model whose design and code will not be made public, which means developers cannot treat it like a downloadable base model in the way they treated Llama. (bloomberg.com) Meta is also tying the model directly to the social products that already have billions of users and years of behavioral data. In its launch post, Meta said Muse Spark will unlock features that can cite recommendations and content people share across Instagram, Facebook, and Threads. (about.fb.com) That gives Meta something an open model cannot guarantee on its own: control over the full route from model to user. If the assistant lives inside WhatsApp chats, Instagram search boxes, Facebook feeds, Messenger threads, and glasses, Meta can ship one model update across its own surfaces without waiting for outside developers to adopt it. (about.fb.com) The timing also tells you what problem Meta thinks it is solving. CNBC reported that Muse Spark is the first major model led by Alexandr Wang at Meta Superintelligence Labs, and TechCrunch said the lab was created after Mark Zuckerberg grew unhappy with Meta’s pace and with Llama lagging behind OpenAI and Anthropic. (cnbc.com, techcrunch.com) Meta is still saying bigger Muse models are coming, but the first release is aimed at product distribution, not community adoption. The company called Muse Spark “an early data point” with larger models in development, while keeping the current release inside Meta AI and a limited partner preview. (about.fb.com) If that pattern spreads, the artificial intelligence market starts to look less like app stores built on shared foundations and more like cable bundles where each giant company keeps the best channel for itself. Meta spent 2023 through 2025 teaching developers to think in terms of Llama; in April 2026, it started teaching them that the most important Meta model may now live behind Meta’s own doors. (ai.meta.com, forbes.com, about.fb.com, bloomberg.com)