Meta abandons Llama for Muse Spark

Meta’s AI strategy just got a lot less open. For three years, Llama made Meta look like the big company willing to give developers actual model weights instead of just an API. But the company’s latest moves point somewhere else. Muse Spark launched on April 8 as the first model from Meta Superintelligence Labs, and by Meta’s April 29 earnings report it was already being framed as the engine behind new business and consumer AI products. (about.fb.com) ### What actually changed? The big shift is not one press release saying “Llama is dead.” Meta has not said that in those words. The shift is structural. Muse Spark is the new flagship model family, it was rebuilt from scratch over nine months inside Meta Superintelligence Labs, and Meta is rolling it across Meta AI, WhatsApp, Instagram(about.fb.com)— current models remain available, but Meta has not clearly committed to frontier-level future Llama development. (about.fb.com) ### Why does Muse Spark feel different from Llama? Because the product model is different. Llama was useful partly because developers could download weights, self-host, fine-tune, and build around it. Muse Spark is the opposite shape. Meta says it powers the Meta AI app and website now, with API access only in private preview for select(about.fb.com)able weights. That makes Muse Spark less like a toolkit and more like a utility pipe controlled by Meta. (about.fb.com) ### Why did Meta make this turn? Basically, Llama stopped looking like a winning frontier strategy. CNBC says Muse Spark arrived after Meta spent heavily to bring in Alexandr Wang and rebuild its AI effort following the disappointing reception to Llama 4. Meta’s own blog leans into that reset language — “rebuilt our AI stack from the gr(about.fb.com)to go. (cnbc.com) ### Is this just about bragging rights? No — it’s about monetization. On the earnings call, Meta said its business AI tools were already facilitating about 10 million conversations per week as of late March, up from 1 million at the beginning of the year. Zuckerberg also said those tools are free for most bus(cnbc.com)duced as a gift to the ecosystem. It is being wired into products Meta can charge through. (techcrunch.com) ### Where does capex fit in? Meta is spending like a company that thinks AI infrastructure is the business, not just a feature. Its April 29 earnings release showed $19.84 billion in quarterly capital expenditures. The same release called out Muse Spark as the first model from Meta Superintelligence Labs in what(techcrunch.com)ayer will pay back through ads, business messaging, subscriptions, or APIs. (investor.atmeta.com) ### What does this mean for developers? If you built on Llama because you wanted control, Muse Spark is not a drop-in successor. That’s the catch. You are choosing between staying on older Llama-family workflows, waiting to see whether Meta releases any genuinely open future models, or moving t(investor.atmeta.com)the company that sells them, but worse for people who want to open the case. (thenewstack.io) ### So is Llama over? Not formally. But as a strategic center of gravity, it looks sidelined. Meta’s public messaging, product rollout, and business metrics all now point to Muse Spark. That does not erase Llama’s installed base or its importance to the last wave of open-weight AI. But it does mean developers should stop assuming Meta’s main AI ambition is to keep serving as open AI’s biggest patron. (about.fb.com) ### Bottom line? The news is not just that Meta launched a new model. It’s that Meta appears to have moved from “here are the weights” to “come through our interface.” For developers, that is a business-model change disguised as a model launch. (about.fb.com)