Meta's Llama 4 Hits a Billion Downloads and Open Source AI Has a New Conversation to Have

Meta hit a milestone with its Llama model family this summer that would have seemed almost absurd to predict two years ago. The Llama 4 series has passed one billion downloads across Hugging Face and Meta's own distribution channels, making it almost certainly the most widely deployed large language model family in history by that measure. For context, that is more than three times the download count Llama 2 reached in its entire lifecycle.

The numbers matter because they tell you something real about where open source AI has landed. Llama 4 is not just popular with tinkerers and researchers. It is running in production at thousands of companies, embedded in enterprise software stacks, powering customer service tools, coding assistants, and internal knowledge bases. The download count is a proxy for deployment breadth, and the breadth is now significant.

Meta released Llama 4 in several sizes earlier this year, with the Scout and Maverick variants designed for different use cases. Scout is a smaller, faster model optimised for on-device and edge deployment. Maverick is the full-scale version that competes more directly with OpenAI's GPT-4o and Anthropic's Claude Sonnet on benchmarks. A third variant, Llama 4 Behemoth, which Meta described as still in training at launch, is the one aimed at the frontier.

What the Benchmark Picture Looks Like

Llama 4 Maverick performs competitively on most standard benchmarks, which is a stronger statement than it sounds given where Llama 2 sat relative to commercial models. On MMLU, the broad knowledge benchmark, Maverick scores around 85 percent, which puts it in the same neighbourhood as GPT-4o. On coding benchmarks like HumanEval, it sits a little behind the very best closed models but ahead of most open alternatives.

The more interesting comparison is practical performance per compute dollar. Because Llama 4 models can be self-hosted or run on inference providers with openly competitive pricing, the cost of generating tokens is substantially lower than going through OpenAI or Anthropic's APIs for high-volume applications. For a company doing tens of millions of API calls per month, the difference in cost is not marginal.

This is increasingly the argument Meta is making. Not that Llama 4 is the smartest model available, but that it is smart enough for most production tasks, and cheaper to run at scale. It is a compelling pitch, and clearly a lot of organisations are buying it.

The Open Source Label Gets Complicated

It is worth being clear about what open source means in this context, because the term is doing a lot of work. Llama 4 is available under a custom licence that allows commercial use for companies below a certain user threshold. Above 700 million monthly active users, you need a separate agreement with Meta. That covers essentially every major platform, which means the very largest potential users are not actually getting the same thing as smaller developers.

The Open Source Initiative has not certified the Llama licence as open source because the weights are available without the full training data and code, and the use restrictions apply. This is a genuine debate in the AI community and not a pedantic one. When people say Llama 4 is open source, they usually mean open weights, and that is meaningfully different.

None of this has dampened adoption, clearly. But it does matter for the longer-term question of what kind of AI infrastructure the world is building. A billion downloads of a model that is open to most but not all users, and whose training process is not fully reproducible, is a different thing from genuinely open infrastructure.

What Comes Next

Meta has been consistent in its position that open models are strategically good for the company because they commoditise AI capabilities that Meta then builds its own products on top of, while also building goodwill with the developer community that creates the ecosystem. It is a smart strategy even if it is not pure altruism.

Llama 4 Behemoth, when it arrives, will be the real test of whether Meta can compete at the frontier. The Scout and Maverick models are competitive but not leading. If Behemoth genuinely pushes the state of the art with an open weights release, that would be a different kind of moment. The developer community is watching.