Most AI labs guard their best models like trade secrets. Meta does the opposite. It builds frontier-scale language models, then publishes the weights for anyone to download and run. The Meta Llama family has now passed 1.2 billion downloads, a number Mark Zuckerberg announced at the company's first LlamaCon developer event in April 2025. For a company that spends tens of billions of dollars a year on AI infrastructure, giving the output away looks strange until you look at what Meta actually gets back.

This is not charity, and it is not quite open source either. It is a calculated bet on where value accrues in the AI stack. Understanding that bet tells you a lot about how the next few years of the market will play out.

What Meta Llama Actually Is

Llama is a family of large language models. The first version leaked to the public in early 2023; since then Meta has shipped Llama 2, Llama 3, and in April 2025, Llama 4. The Llama 4 release moved to a mixture-of-experts design, where only a slice of the model activates for any given token. Two models shipped as open weights: Scout, with 17 billion active parameters across 16 experts and a very large context window, and Maverick, with the same active-parameter count spread across 128 experts. A larger model, Behemoth, was previewed but held back.

The word that matters is "weights." When Meta releases a model, it publishes the trained parameters, the numerical guts that let you run the model on your own hardware. You accept a license, download the files from Hugging Face or llama.com, and you are off. You do not need to call Meta's servers, pay per token, or send your data anywhere. That is the core difference from a closed API like OpenAI's or Anthropic's, where the model stays behind a paywall you rent access to.

Open Weights Is Not Open Source

Meta markets Llama as open source. It is not, and the distinction is more than pedantic. Open weights means you get the finished model. Open source, in the traditional software sense, would mean you also get the training data and the full recipe to rebuild it, plus the freedom to use it for any purpose. Llama gives you neither the data nor unrestricted use.

The Llama Community License carries real limits. The headline one: if your product has more than 700 million monthly active users, you cannot just use Llama, you must request a separate license that Meta can grant or refuse at its discretion. That clause is aimed squarely at Meta's rivals, the handful of companies large enough to turn Llama into a competing consumer product. The license also restricts using Llama to train competing models and, in newer versions, carves out some multimodal use in the EU. The Open Source Initiative has stated plainly that the Llama license is not open source, because it discriminates against certain users and fields of use.

So the accurate label is open weight with strings attached. For the vast majority of developers and companies, those strings never bind. For the three or four companies that could rival Meta at scale, they bind hard. That asymmetry is the whole point.

Why Meta Gives It Away

Strip out the idealism and the logic is commercial. Meta does not sell AI models, so it does not lose revenue by releasing them. What it sells is advertising against attention, and AI that makes its apps stickier and its ad targeting sharper feeds that machine directly. Everything else is strategic positioning.

First, releasing weights commoditizes the layer Meta's competitors want to charge for. If a capable model is free, it is harder for OpenAI or Google to build a durable business selling model access alone. Meta is happy to burn down a market it does not compete in, the way Google once open-sourced Android to keep any single rival from owning mobile.

Second, openness buys a developer ecosystem. When Meta hit 1 billion Llama downloads in March 2025, it also gained a global base of engineers who build tools, fine-tunes, and infrastructure around its models for free. That community hardens Llama as a default, surfaces bugs, and produces improvements Meta can learn from. Talent follows too; researchers prefer to work where their output ships publicly.

Third, it is a hedge against dependence. Meta does not want to rent its AI future from a competitor. Owning an open model it controls the release cadence of means it never negotiates from weakness.

What the Open-vs-Closed Split Means for the Market

The market is settling into two camps, and Meta Llama anchors the open one alongside models from Mistral, DeepSeek, and others. Closed frontier labs still tend to lead on raw capability at the very top end. Open weights lead on control, cost, and privacy.

For founders and operators, the practical read is this. If you need the absolute best reasoning and are fine renting it, closed APIs are simple and strong. If you care about running models on your own infrastructure, keeping data in-house, avoiding per-token bills at scale, or fine-tuning deeply, open weights like Llama are the obvious base. Many teams now run both, a closed model for the hardest tasks and an open one for the high-volume, cost-sensitive work.

The strategic risk to watch is that "open" here depends on Meta's continued willingness to publish. The license can change between versions, and it already has. Building on Llama means building on a gift that comes with terms Meta writes. That is a very different foundation than a permissive open-source project no single company can revoke. Useful, powerful, and free, but not neutral.

Share this article: