Meta Launches Llama 2, Its Free Open AI Model

Meta announced today the launch of Llama 2, the second generation of its family of open language models. The headline change isn't just technical: for the first time, Meta is allowing free commercial use of these models, with one explicit exception written into the license — companies with more than 700 million monthly active users must request special permission from Meta before using it, a clause aimed squarely at its biggest tech rivals.

The launch comes with a heavyweight partner: Microsoft. Llama 2 will be available preferentially on Azure, Microsoft's cloud platform, and the company has also worked to optimize the model to run efficiently on Windows. It's a striking twist: Microsoft is OpenAI's biggest investor and the backer behind GPT-4, and it's now also supporting distribution of an open model that competes directly with that ecosystem.

What's different from Llama 1

Meta released the first version of Llama back in February of this year, but access was restricted to researchers and came with no commercial license — which didn't stop the model's weights from leaking and spreading widely across the internet anyway. Llama 2 fixes that starting point: it's being released directly under an open, commercial license, with no need for leaks or informal access.

The family comes in three sizes based on parameter count — 7 billion, 13 billion and 70 billion — letting developers choose between lighter models that run on modest hardware and more powerful versions for demanding tasks. According to Meta, Llama 2 was trained on 2 trillion tokens, 40% more data than its predecessor, and doubles the context window to 4,096 tokens — the amount of text the model can factor in at once when generating a response.

Alongside the base models, Meta is also releasing Llama 2-Chat, a version fine-tuned specifically for conversation, trained using reinforcement learning from human feedback (RLHF) — the same technique ChatGPT popularized to make its responses more useful and less likely to produce problematic content.

Why the free license matters

Until now, anyone wanting to build a commercial product on top of a powerful language model had essentially two options: pay for API access, as with OpenAI or Anthropic, or settle for open models with weaker capabilities or more restrictive usage terms. Llama 2 changes that math. A startup, a university or a mid-sized company can now download the model's weights, run it on its own infrastructure and build a product on top of it without paying Meta any licensing fees — something no model of this capability level has offered until now.

Meta has made the model available to developers through several channels: besides Azure, it will be available on Hugging Face, the go-to platform for sharing AI models, and on Amazon Web Services, among other cloud providers. That multi-platform availability means adoption won't hinge on a single provider.

The 700-million-user threshold

The clause covering companies with more than 700 million monthly users deserves a closer look. In practice, it shuts a handful of companies out of the free license: the giants with mass-consumer products that could otherwise use Llama 2 to compete directly with Meta on its own turf. It's a way of opening the model up to the broader developer ecosystem, startups and mid-sized companies, while reserving the right to negotiate separately with those with enough scale to become direct rivals.

An open race, with caveats

Meta says that in its own helpfulness and safety testing, Llama 2-Chat matches or beats other open-source models available today, though the company acknowledges it still trails the most advanced closed models, such as GPT-4. That gap matters less than it might seem for the broader ecosystem: until now, anyone wanting a powerful open model had to rely on leaked versions or projects with restrictive commercial licenses. Llama 2 offers, for the first time, a combination of scale, competitive performance and freedom of use with no direct precedent.

Meta's move also carries a broader strategic logic. By giving away the technology, the company forgoes charging directly for the model, but gains something else instead: tens of thousands of developers building on its infrastructure, generating improvements, finding bugs and expanding the Llama ecosystem — all without Meta having to bear the cost of maintaining its own commercial API. It's the same calculation that led Google to open up Android more than a decade ago, now applied to language models.

For developers and businesses, the immediate effect is having a real alternative to paid APIs, with control over their own data and no dependence on an external provider for every query. For the rest of the industry, the open question is how long it will take Meta's competitors to respond with openings of their own, and whether OpenAI's and Google's edge in closed models can remain sustainable against an open ecosystem that's now moving with resources of this magnitude.