Explore AI
Open weights are not open source
The distinction sounds pedantic until you try to sell something you built with a model you were told was open.
Almost every article about local AI uses open source to mean you can download the file. That is not what the words mean, and the difference has legal consequences.
Three different things, usually confused
Open weights — the trained parameters are downloadable. You can run them, fine-tune them, often redistribute them.
Open source — the weights, and the training code, and the data, under a licence that permits any use. Almost nothing in AI meets this.
Free — costs nothing to use. Says nothing about either of the above. A closed model can be free; an open-weights model can be expensive to run.
The licence spectrum
| What the card says | What it usually means |
|---|---|
| Apache-2.0, MIT | Genuinely permissive. Use it commercially, modify it, ship it. This is what people think "open source" means. |
| A bespoke community licence | You may use it, with conditions: a clause about very large-scale commercial use, a naming requirement, or a right to change the terms later. |
| Non-commercial | Research and personal use only. You cannot build a product on it. |
| OpenRAIL-style | Permissive with use restrictions — you must pass the same restrictions downstream. |
| "Other" or unlabelled | Read the actual text. This label sits on some of the most-downloaded models on the hub. |
The model catalog carries a commercial-use column for exactly this reason, and the families show how much the terms vary even inside one family.
Why it matters in practice
You may not be allowed to sell what you build. This is the one that bites. A weekend project on a non-commercial licence is fine; the same code in a product is a legal problem.
Terms can change. A community licence is not a promise. New versions arrive with different conditions, and the model you built on stays where it is while the family moves on.
You cannot audit it. With weights alone you can inspect behaviour, not provenance. You cannot verify what it was trained on, or reproduce it.
You depend on the publisher. Open weights are a gift, not an obligation. It can be withdrawn, or simply never updated.
The honest summary
Open weights have been enormously good for users. They are why you can run a capable model on a laptop at all, why prices across the industry have fallen, and why no single company can switch local AI off.
But "open source" is mostly marketing. A handful of organisations are genuinely open — the OLMo family releases weights, data and training code under a permissive licence, and is roughly alone in doing so. Most of what you will see described as open source is downloadable weights under terms no software licence would recognise.
How to check in thirty seconds
1. Open the model's page on the hub. 2. Find the licence tag — every model card has one. 3. If it is not Apache-2.0 or MIT, read the actual text before building anything you intend to sell. 4. If there is no licence at all, treat the model as unusable for commercial work.
Thirty seconds, against the possibility of rebuilding a product later.