Meta ships Muse Spark
Meta just released a new flagship model called Muse Spark and is rolling it into the Meta AI assistant to power product features rather than just publish research. This is the first model from the superintelligence team Meta built after hiring Alexandr Wang, and the company ties the model directly to user-facing assistant features rather than keeping it purely internal or research-only. That matters because it shows Meta shifting from capex-and-open-source rhetoric to shipping foundation models that sit behind real product surfaces. (reuters.com)(x.com)
Meta spent most of 2025 talking about giant artificial intelligence spending and open-weight models like Llama. On April 8, 2026, it did something more concrete: it put a new flagship model called Muse Spark directly inside the Meta AI assistant on the web and in the Meta AI app. (reuters.com) (about.fb.com) This is the first public model from Meta Superintelligence Labs, the group Mark Zuckerberg built after a recruiting spree and a $14.3 billion investment in Scale AI that brought in Scale co-founder Alexandr Wang. Reuters said the lab was assembled through a costly talent war and internal restructuring after Meta fell behind rivals in the race to build top-tier models. (reuters.com) (cnbc.com) Meta says Muse Spark was built over nine months after the company “rebuilt” its artificial intelligence stack from the ground up. In plain English, that means Meta is saying this is not just another tune-up to Llama, but a new base engine for the assistant people actually use. (about.fb.com) (techcrunch.com) The model is not being thrown open to everyone the way Meta handled earlier Llama releases. Reuters reported that Meta is keeping Muse Spark mostly inside its own products, with only a private application programming interface preview for selected partners instead of a broad public release. (reuters.com) (about.fb.com) That product-first approach shows up in the rollout list. Meta says Muse Spark powers the Meta AI app and website now, and it plans to push the model into WhatsApp, Instagram, Facebook, Messenger, and its artificial intelligence glasses in the coming weeks. (about.fb.com 1) (about.fb.com 2) Meta is pitching Muse Spark as a multimodal model, which means one system can handle text, images, and voice instead of passing work between separate tools. The company says the model is aimed at assistant tasks like conversation, reasoning, visual understanding, and coding help rather than just benchmark demos. (about.fb.com) (x.com) Wang made that tradeoff explicit on X when he said Meta is using Muse Spark to power Meta AI services now and is already working on larger models next. That is a different message from “here is a research model,” because the test is no longer a leaderboard screenshot but whether hundreds of millions of people keep using the assistant. (x.com) (reuters.com) The backdrop is that Meta’s previous artificial intelligence push had started to look uneven. Reuters reported that the company’s earlier models drew criticism, and CNBC said Muse Spark is Meta’s first major large language model since Wang arrived to reset the effort. (reuters.com) (cnbc.com) So the real story is not just that Meta named a new model on April 8. It is that after spending billions on chips, talent, and data, Meta has started tying its newest foundation model to visible consumer surfaces where people can tell within minutes whether it feels smarter than what came before. (reuters.com) (about.fb.com)
