Buried in Meta’s license for Llama is a clause that says any company with more than 700 million monthly active users cannot use the model freely and must request a separate license from Meta first. That single number, 700 million, tells you almost everything about what “open” means in AI today. A truly open project does not carve out exceptions for whichever companies happen to be large enough to compete with the one that built it. Yet Llama is described in headline after headline as an open source model, next to Qwen, Mistral, DeepSeek, and a dozen others that get the same label despite running on very different terms.

what open source was supposed to mean

The term comes from software, where the Open Source Initiative (OSI) has a specific definition: the source is available, and so is everything needed to inspect, modify, and rebuild the thing from scratch, with no restriction on who gets to use it or for what. Applied to a language model, that would mean the weights, the training code, and the training data itself are all available, so that in principle anyone with enough compute could reproduce the model from raw material. Almost nothing released under an AI company’s banner meets that bar. The training data in particular, the exact mixture of web crawl, licensed text, and curated sets described in Training Data: Where It Comes From, almost never ships alongside the weights. Companies treat that mixture as the one part of the recipe worth protecting.

what actually gets released, and under which license

What ships instead is “open weight”: the finished, trained parameters, downloadable and runnable, with the data and pipeline that produced them kept private. The license attached to those weights varies more than people assume. Qwen, Mistral, and Gemma 4 use Apache 2.0, a genuinely permissive software license that imposes almost no restrictions on commercial use or redistribution. DeepSeek, GLM, and Phi-4 ship under plain MIT, equally permissive. Kimi K2 uses a modified MIT variant that adds a few extra conditions on top of the standard text. Llama sits in a different category entirely: a custom, community-style license that reads like an open license for the vast majority of users and then quietly excludes the handful of companies large enough to matter competitively, via that 700 million monthly active user line. Four models, four license philosophies, and only two of them (Apache 2.0 and MIT) are licenses the OSI would recognize as open source in the software sense, even though none of the four release their training data.

It helps to think of open weight releases the way you’d think of buying a car with the hood welded shut. You can drive it anywhere, put your own tires on, repaint it, take it apart at the body panels, and hand it to a mechanic for a full tune. What you cannot do is open the engine block and see what alloy it’s made of or how it was cast, because that part was never meant to come apart. The car works, and works well. It is just not the same thing as owning the factory.

That distinction matters in practice, not just in licensing footnotes. Being able to download and run weights, fine-tune them on your own data, and inspect how the model behaves on inputs you control is real, useful access, and it is the reason open-weight models have become competitive with closed ones on many tasks. But in the large majority of cases you still cannot reconstruct the model from scratch, and you cannot know with certainty what it was trained on or in what proportions. The word “open” in front of a model’s name is doing more marketing work than technical work. What it actually promises, almost every time, is a car you can drive, not a factory you can inspect. The next piece in this series, on how parameter counts are actually built, looks at another place where a single number on a model’s spec sheet hides more than it reveals: the difference between dense and mixture-of-experts architectures.