For three years, Mark Zuckerberg positioned Meta as the industry’s champion of open AI — publishing manifestos about it, criticizing OpenAI publicly for going proprietary, and building Llama into the most downloaded AI model family on the planet, reaching 1.2 billion cumulative downloads by early 2026. On April 8, 2026, Meta shipped its first fully closed-source frontier model, Muse Spark, from its newly formed Meta Superintelligence Labs. The open-source era at Meta was over. And if you’d been paying close attention to the actual license terms rather than the marketing, you’d have known it was never quite as open as advertised in the first place.
The Word “Open” Was Doing a Lot of Work
Here’s the uncomfortable detail underneath this whole saga: the Open Source Initiative — the organization that literally wrote the industry’s definition of what “open source” means — has been saying publicly since at least 2024 that Meta’s Llama license doesn’t qualify as open source at all. Not a minor technicality. OSI’s assessment found the license fails multiple points of the actual Open Source Definition: it fails “freedom 0,” the basic freedom to use software for any purpose, and it violates the Definition’s explicit prohibitions against discriminating against users or restricting fields of use.
The specific mechanism is a clause with no real equivalent in genuine open-source licensing: Section 2 of the Llama Community License requires anyone whose product or service crosses 700 million monthly active users to request a separate commercial license directly from Meta — one Meta can grant or deny entirely “in its sole discretion.” For most startups that threshold is irrelevant. But the clause itself violates a core open-source principle: that a license can’t discriminate against specific users or restrict who’s allowed to build on the technology, regardless of how large or small that group turns out to be. Newer versions of the license added further restrictions too, including barring anyone located in the EU from using the model at all, without explanation.
Meta wasn’t the only offender, either. Google’s Gemma Terms of Use and other “open-weight” model licenses carry their own versions of the same problem — acceptable-use policies restricting entire product categories, attribution requirements with no equivalent in standard open-source terms, and in some cases the ability for the model provider to revoke access after the fact. The Open Source Initiative published its official Open Source AI Definition (OSAID) v1.0 in October 2024 specifically to draw a clear line here — requiring genuine access to training data information, model architecture, and training code — and by that standard, most of the models the industry casually calls “open,” including Llama, Gemma, and Qwen, simply don’t qualify.
Why This Wasn’t Just a Legal Technicality
The gap between the marketing and the license terms created real, practical risk for companies building on these models — not hypothetical risk, but the kind that shows up in actual legal guidance now circulating for startups. When an engineering team pulls Llama down from Hugging Face or drops Gemma into a production pipeline, the license file is usually the last thing anyone reads closely. But those licenses contain materially different obligations than a standard open-source license: acceptable-use policies that can restrict entire categories of product, branding and attribution requirements, and revenue or user thresholds that can trigger the need for a separate commercial agreement with the model’s creator, on terms that creator controls unilaterally.
The Open Source Initiative’s head of community put the core problem plainly: the restrictive and inconsistent licensing of so-called “open” AI models creates real uncertainty for commercial adoption, and while these models are marketed as open, their actual terms impose legal and practical hurdles that discourage businesses from integrating them into real products. To be fair to some of the model providers here, not every company has hidden its intentions — Cohere, for instance, has been explicit that its models support scientific research but not unrestricted commercial use, which is a materially more honest framing than calling a restricted license “open source” outright.
Then Meta Just… Stopped
The April 2026 shift to Muse Spark didn’t happen in isolation, and the reasoning behind it is more interesting than “Meta changed its mind.” Three converging factors reportedly drove the decision. First, Llama 4 faced a genuine benchmark-gaming controversy that damaged the credibility open weights had built for Meta in the AI community — once trust in Llama’s benchmark claims eroded, much of the strategic value of releasing it openly went with it. Second, Meta’s $14.3 billion investment in Scale AI created a proprietary data pipeline that’s arguably better protected behind closed weights than exposed through an open release. Third, and reportedly the least-discussed but most decisive factor: adversarial distillation, where competitors take a company’s open weights and use them to train cheaper knockoff models that compete directly with the original — a dynamic that makes releasing a genuinely frontier-tier model as open weights strategically costly rather than strategically generous.
The competitive backdrop makes that calculus even sharper. As of April 2026, U.S. models accounted for roughly 35% of global Llama deployments, while Chinese labs like Alibaba and DeepSeek had grown to account for around 41% of downloads on platforms like Hugging Face by late 2025 — meaning a meaningful share of the benefit from Meta’s years of open releases was flowing to competitors building on top of that same openness, not necessarily strengthening Meta’s own competitive position.
Open Source AI Isn’t Dead — It’s Relocating
The more useful way to read this moment isn’t “open-source AI is over.” It’s that the center of gravity is shifting away from the company that spent three years claiming to lead it. As of April 2026, five of six major open-source model families were built on mixture-of-experts architecture and shipped under genuinely permissive licenses — DeepSeek’s V4 family among them, released under an actual MIT license, which is the real open-source standard Llama’s license was never quite willing to meet. DeepSeek V4-Pro’s technical efficiency claims are notable on their own merits — a hybrid attention design that reportedly lets it run a 1-million-token context using a fraction of the compute and memory of the previous generation.
That said, “genuinely open license” doesn’t automatically mean “no strings attached” in a different sense: Anthropic publicly accused DeepSeek in February 2026 of using fraudulent accounts to generate large volumes of Claude conversations, reportedly for training purposes — a reminder that license permissiveness and responsible practice are two entirely separate questions, and a genuinely open license doesn’t by itself vouch for how a company got there.
What This Actually Means If You’re Building on Any of These Models
The practical lesson for developers and companies isn’t “avoid open-weight models.” It’s that the word “open” in AI licensing has stopped functioning as a reliable signal, and reading the actual license — not the marketing page, not the Hugging Face model card summary — has become a genuine engineering and legal prerequisite rather than an afterthought. A model can be free to download, technically capable, and still carry meaningful legal exposure buried in a license clause nobody on the team reads until it matters.
Meta’s exit from open weights is, in a strange way, a moment of clarity rather than a loss: it removes the largest, most visible source of the “open source” confusion that’s defined this space for the past three years. What replaces it — DeepSeek, Qwen, and a genuinely multipolar field of model providers with varying degrees of real openness — at least forces the question that Llama’s marketing always let people avoid: is this actually open, or does it just look that way from a distance?