Meta drops Llama for Muse Spark

Meta’s AI story just got messier than the headline suggests. The simple version is not “Llama is dead.” The real shift is that Meta’s newest flagship model, Muse Spark, is closed at launch, while Llama stays around as the older open-weight line. That matters because a lot of developers treated Llama as proof that Meta would keep pushing frontier-ish models in an open direction. Turns out the company’s center of gravity moved. (thenewstack.io) ### Did Meta actually drop Llama? Not exactly. Meta has said current Llama models will remain available, which is very different from saying Llama remains the company’s main bet. The company’s April push went into Muse Spark, the first major model from Meta Superintelligence Labs, the group led by Alexandr Wang after Meta’s 2025 AI reorganization. So Llama looks less canceled than depr(thenewstack.io)ort is going. (thenewstack.io) ### What is Muse Spark? Muse Spark is Meta’s new multimodal reasoning model. It takes text, images, and speech, supports tool use, and runs different reasoning modes, including a multi-agent “contemplating” mode. Meta is positioning it as the model underneath Meta AI across its apps and devices, including Ray-Ban glasses. That alone tells you the goal changed — this is not just a model(thenewstack.io)ta’s consumer stack. (deeplearning.ai) ### Why does the closed launch matter so much? Because the gap is not philosophical. It is operational. Llama’s appeal was that teams could download weights, self-host, fine-tune, and control deployment. Muse Spark does not offer that. At launch it is available through Meta AI, with API preview only for selected partners. If y(deeplearning.ai) are not moving versions. You are changing business model and infrastructure. (thenewstack.io) ### Is Muse Spark actually better? In some areas, yes. Meta says Muse Spark performs competitively on reasoning, multimodal work, and health-related benchmarks, and the company claims much better training efficiency than Llama 4 Maverick — roughly similar capability with more than an order of magnitude less training processing. But Meta also admits Muse Spark still trails in some codin(thenewstack.io)placement. It is a new architecture optimized for Meta’s own priorities. (deeplearning.ai) ### Why would Meta pivot now? Basically — control and product fit. Open-weight models are great for ecosystem goodwill, but they do not lock developers into Meta’s own surfaces. A closed model tied to Meta AI, apps, and glasses does. There is also a safety and governance angle. Meta’s preparedness report says Muse Spark showed(deeplearning.ai)ed safeguards inside Meta’s deployment stack. Closed delivery makes those controls easier to enforce. (ai.meta.com) ### So what should developers take from this? Do not confuse “available today” with “strategic priority tomorrow.” That is the lesson. Even when a company has an open-source reputation, the flagship path can still swing back to proprietary once the business case changes. Meta may still release some future models more openly — that possibility is still floating(ai.meta.com)ama-shaped. (thenewstack.io) ### What’s the bottom line? This is really a warning about dependency. If you built on Llama, you still have Llama. But if you built on Meta’s long-term openness, that assumption just got a lot shakier. (thenewstack.io)