Most of what you are told is open AI is not open source
The most popular image model on the internet cannot legally be used to make something you sell. Nobody mentions this.
At a glance
- FLUX.1 dev
- non-commercial licence
- XTTS-v2
- non-commercial licence
- Stable Diffusion XL
- commercial, with use restrictions
- OLMo
- weights, training data and code all released
Downloadable is not the same as open. Free is not the same as free to sell. And the gap between those things is where a weekend project becomes a legal problem.
Take a tour of the most popular open-weight models, checking only the licence field.
The ones that are not what you think
FLUX.1 dev — one of the most widely used image models, described as open everywhere. Its licence forbids commercial use. You cannot build a product on it. You cannot sell what it makes.
XTTS-v2 — a popular voice-cloning model, downloaded millions of times. Licensed for non-commercial use only.
Stable Diffusion XL — the most-used image model ever released. Commercial use is permitted, with restrictions attached that you are required to pass downstream.
Llama — openly downloadable, and under a community licence with conditions, including limits on very large-scale commercial deployment. It is not an open-source licence, and Meta can change the terms between versions.
MiniMax — several releases carry bespoke or non-commercial terms.
Why nobody notices
Because the question people ask is "can I download it?" — and the answer, satisfyingly, is always yes.
The question that matters is "can I sell what I build with it?", and it appears in a licence file that most tutorials never open.
There is a second reason. The words open source carry a lot of goodwill, built over decades by software that genuinely earned it. Applying them to downloadable model weights borrows that goodwill without the obligations that normally come with it.
The one family that does it properly
The OLMo models, from the Allen Institute for AI, release the weights, the training data, and the code that produced them, under a permissive licence.
That is what open source means. Almost nothing else in AI qualifies — and OLMo is not the most capable model available, which tells you something about why.
Thirty seconds that saves you a rebuild
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 you build anything you intend to sell. 4. If there is no licence, treat the model as unusable for commercial work.
Thirty seconds, against the possibility of rebuilding a product later.
The honest summary
Open weights have been enormously good for users. They are the reason you can run a capable model on a laptop, and the reason prices across the industry have fallen.
But if you are building something to sell, assume nothing is open until you have read the licence — and check again when the next version arrives, because the terms can move.
Every claim in this file is checked against primary sources — how we verify. Spotted an error? Tell us.
Get the next file before it's public
One story a week, chosen from the archive. The kind you repeat at dinner and nobody believes.
One dispatch a week. Unsubscribe anytime.
More Technology
AllOpenAI publishes open weights now
The company that spent years arguing against open models now ships them under the most permissive licence there is.
Who Actually Invented "Artificial Intelligence"?
The phrase was chosen in 1955 partly to avoid an argument with a rival researcher — and its inventor spent decades watching people stop calling things AI as soon as they worked.