Meta’s most important AI shift is not Muse Spark itself. It is Alexandr Wang saying plainly that the old Llama playbook does not hold at the frontier.
For years, Meta used open models as a weapon against better funded rivals. Llama gave researchers, startups, and developers something they could download, alter, and run without waiting for permission from OpenAI, Google, or Anthropic. Now the executive Mark Zuckerberg hired to fix Meta’s AI effort is saying the company hit a limit.
In a Bloomberg Tech interview reported today by the Times of India, Meta chief AI officer Alexandr Wang said Muse Spark stayed closed because the company’s internal safety testing found risks it could not contain if the model weights were released publicly. That is not a small product choice. It is the clearest admission yet that Meta’s open AI posture is being rewritten by the same frontier model race it once tried to disrupt.
Muse Spark arrived on April 8 as the first model from Meta Superintelligence Labs, the group built after Meta’s reported $14.3 billion investment in Scale AI and Wang’s move from Scale into Zuckerberg’s inner circle. Business Insider reported at the time that the model powers the Meta AI app and meta.ai, includes a contemplating mode for coordinating multiple agents, and was built with health responses shaped in consultation with more than 1,000 physicians. Meta also said coding remained an area with performance gaps.
That last detail matters because Muse Spark is not simply a safety story. It is also a competitiveness story. Wired reported in April that Artificial Analysis placed Muse Spark within the top five models it had benchmarked, while the Financial Times reported last week that some Meta staff still preferred Anthropic’s Claude for software development tasks. A closed model that still trails in a key enterprise use case is a harder sell than an open model that startups can forgive, modify, and champion.
Meta’s own safety paper gives Wang’s comments more weight. The Muse Spark Safety and Preparedness Report, published on arXiv on May 14, said evaluations covered chemical and biological, cybersecurity, and loss of control risks under Meta’s Advanced AI Scaling Framework. The report said Muse Spark’s deployment inside Meta AI carried acceptable residual risk after safeguards, but it also said chemical and biological capabilities were likely in the high risk category before mitigations.
That is the part open model advocates cannot hand-wave away. Once weights are public, a company can publish guidance and usage policies, but it cannot reliably pull the model back from every server, lab, hobby project, and malicious fork. Meta can monitor a model inside Meta AI. It cannot monitor every copy of a frontier system once the copy exists outside its walls.
This is why Wang’s Bloomberg comments land differently from the usual corporate caution around AI safety. Meta has spent years gaining goodwill from developers by releasing Llama models more openly than its biggest American rivals. The company was never as open as the pure open source label suggested, since Llama carried license restrictions and did not disclose the full training recipe, but it still gave builders more direct control than a closed API. Muse Spark breaks that habit at the point where the model is meant to be most strategically important.
Open AI may be moving down market
The awkward possibility for startups is that open AI becomes strongest just below the frontier. Smaller labs, enterprise teams, and developer communities may still get very capable open-weight models. They may even get enough performance for most products. But the biggest frontier systems, the ones with agentic behavior, deeper reasoning, stronger multimodal ability, and more dangerous dual-use knowledge, are now drifting toward the same controlled-access model that OpenAI, Anthropic, and Google have used all along.
Meta can afford that turn because its business is not built on selling model access alone. Its advertising machine still throws off the cash needed to hire expensive researchers, buy chips, and absorb the cost of long training runs. The company can put Muse Spark into Facebook, Instagram, WhatsApp, and the standalone Meta AI app, then decide later how much of the system outside developers deserve to touch.
That is a very different bargain from the Llama era. Back then, Meta’s openness helped it recruit a community and pressure rivals. With Muse Spark, safety and control have moved closer to the center of the business model. The company is no longer only asking whether openness helps adoption. It is asking whether openness weakens its ability to manage risk, protect an expensive model, and keep pace with Claude, Gemini, and GPT-class systems.
Wang has said future Meta models may still include open releases, and that caveat is important. The shift is not the death of open AI. It is a narrowing of where open AI fits. The frontier is becoming too expensive, too sensitive, and too tied to platform control for Meta to treat it like a public download by default.
That leaves developers with a less comfortable version of the old Meta promise. They may still get powerful tools from Menlo Park. They should no longer assume they will get the most important ones.
Also read: SEBI is preparing to make AI part of market regulation • AI financing is now large enough to move public markets • KPMG pulled an AI report after its own facts fell apart
