For years Meta's pitch to the AI world was openness. While OpenAI, Google and Anthropic guarded their best systems behind paywalled APIs, Meta released its Llama models for anyone to download, modify and run. That openness won goodwill from developers and kept competitive pressure on rivals. Now Alexandr Wang, the chief AI officer Mark Zuckerberg hired to fix Meta's AI effort, has said plainly what many suspected: at the frontier, the old playbook no longer holds.
The occasion is Muse Spark, the first model from Meta's new Superintelligence Labs and the company's most capable system to date. It shipped in April closed, powering the Meta AI app rather than available as a download. In a Bloomberg interview, Wang explained why: internal safety testing found risks the company could not contain if it released the weights publicly. Meta's own preparedness report, posted to arXiv in May, said Muse Spark's chemical and biological capabilities were likely in the "high risk" category before mitigations were applied.
That is the argument open-model advocates have the hardest time answering. A company can wrap a closed model in usage policies and monitoring, then pull the plug when something goes wrong. Once weights are public, there is no plug to pull. Meta can watch how Muse Spark behaves inside its own app; it cannot watch every copy running on every server, lab and hobbyist fork once those copies exist. For a system with genuine dual-use potential, that asymmetry is the whole ballgame.
Muse Spark is not only a safety story, though, and the two threads reinforce each other. Independent benchmarks placed the model in the top five, yet the Financial Times reported that some Meta staff still prefer Anthropic's Claude for coding, a key enterprise use. A closed model that trails in an important task is a harder sell than an open one that developers will forgive and improve themselves. Keeping it closed removes the single advantage that openness reliably bought.
Meta can afford the shift because, unlike the labs that sell model access, its money comes from advertising. Llama's openness was always partly strategic, a way to recruit a developer community and undercut rivals, and Meta was never as open as the label implied: its licenses carried restrictions and it never disclosed full training recipes. But the business never depended on it. The company can fold Muse Spark into Facebook, Instagram and WhatsApp and decide later, on its own terms, how much outsiders get to touch.
Wang has said future Meta models may still include open releases, and that caveat matters. This is not the death of open AI so much as a narrowing of where it lives. Capable open-weight models will likely keep flowing just below the frontier, good enough for most products. But the most powerful systems, the ones with the deepest reasoning and the most dangerous knowledge, are drifting toward the same controlled-access model their rivals always used. For developers, the old Meta promise survives in weaker form: you will still get powerful tools from Menlo Park, just not the most important ones.