"Open" is one of the most overloaded words in AI. A model gets called "open source," people assume it carries the freedoms of open-source software, and the reality turns out to be narrower. The distinction between open weights and genuine open source is not pedantry — it determines what you are actually allowed to do, and it is at the center of a real fight over the future of the field.
What you actually get
When a lab "opens" a model, what almost always ships is the weights: the trained parameters you can download and run on your own hardware. That alone is enormously valuable — you can run the model privately, fine-tune it, and build on it without calling anyone's API. But weights are the output of training, not the recipe. You typically do not get the training data, the exact training code, or the full procedure needed to reproduce the model from scratch. You get the cake, not the ingredients list.
Why that differs from open source
In software, "open source" has a precise meaning: you get the human-readable source, you can study exactly how it works, and you can rebuild it yourself. Open weights give you something more like a sealed binary you are allowed to run and modify at the edges, but cannot fully reconstruct or audit. You cannot see what data shaped the model's behavior or biases, and you cannot independently verify how it was made. For many uses that is fine; for transparency, reproducibility, and accountability, it is a meaningful gap.
The license catch
There is a second twist: the license. Plenty of popular "open" models ship under custom terms that are not open-source licenses at all. Some restrict commercial use above a certain scale, forbid using outputs to train competing models, or carve out specific applications. A genuine open-source license, by definition, does not discriminate against fields of use. So a model can be freely downloadable and still come with strings that a true open-source project never would. "You can download it" and "you can do whatever you want with it" are not the same claim.
Why labs draw the line where they do
Releasing weights but withholding data and training details is a deliberate middle path. It lets a lab claim the goodwill and ecosystem benefits of openness — developers building on the model, scrutiny that surfaces bugs, mindshare — while protecting the genuinely expensive, competitive assets: the curated data and the training know-how. It can also be a hedge on safety, since full recipes for frontier capabilities are harder to put back in the box once published.
Why the fight matters
This is not a licensing footnote; it is about who controls capability. Open weights are a powerful counterweight to a future where the best models live only behind a handful of corporate APIs — you cannot gate access to something people can download and run themselves. But "open weights" is not the same as a transparent, auditable, freely licensed commons. As governments start vetting access to closed frontier models, the open-weights camp becomes the obvious escape valve. The honest framing keeps that promise grounded.
Why it matters
Before you build on an "open" model, read what you actually received: weights or source, and a real open license or a custom one with limits. The freedom to run a model privately is genuine and valuable. The freedom to study, reproduce, and use it without restriction is a stronger claim — and one fewer models meet than the marketing suggests. Knowing the difference is the difference between a safe foundation and an unpleasant surprise.
Analysis by GenZTech.