Ten licence traps in open AI models
Every one of these is a model you might reasonably pick, that would cost you later. Some carry a licence you cannot sell on; some carry conditions people miss; two are not licence problems at all but size traps, where the number in the model's name describes one part of something much bigger. 3 of the 10 would stop you selling a product outright.
We read the licence file or the model card for each of these and wrote down what it says,
with the date. We did not run most of them — on the shelf here they carry the
REVIEWED label rather than a measurement, and that is deliberate.
This is not legal advice. Every row below links to the source so you can read it
yourself in a minute, which is the only reason this page is worth anything.
The models, one by one
Qwen2.5-VL-3B
The non-commercial licence sits exactly on the small size.
The 7B and the 72B are Apache-2.0. The 3B — the one you reach for first, because it is the one that fits a modest card — is not. So the cheapest model in the family is the one you cannot build a paid product on.
What we did instead: Use the 7B if you can afford the VRAM. If you cannot, this family is not your answer.
check it yourself ↗ · read by us on 2026-09-18
Nanonets-OCR-s
No declared licence, and it is derived from Qwen2.5-VL-3B.
Two problems that compound: there is no licence file to read, and the weights descend from a non-commercial model. A derivative does not escape the licence of what it was made from.
What we did instead: Treat it as non-commercial until someone tells you otherwise in writing.
check it yourself ↗ · read by us on 2026-09-18
moondream3-preview
Forbids paid products that compete with it.
Which covers most of the reasons you would reach for a small vision model commercially. `moondream2` is clean — so check which one you actually pulled, because the names are one character apart.
What we did instead: Use `moondream2`.
check it yourself ↗ · read by us on 2026-09-18
SDXL-Turbo
Commercial use only under 1 million USD of revenue, and you must display 'Powered by Stability AI'.
The revenue cap is the part people notice. The attribution requirement is the part that surprises them later: it goes on the product, not in a file nobody opens.
What we did instead: Fine for a small product if you are comfortable with the badge. Read it again before you grow.
check it yourself ↗ · read by us on 2026-09-18
SD-Turbo
Same Stability licence, same conditions as SDXL-Turbo.
Listed separately because people assume the smaller sibling is freer. It is not: same revenue cap, same attribution.
What we did instead: Same answer as SDXL-Turbo.
check it yourself ↗ · read by us on 2026-09-18
LFM2-VL-1.6B
Commercial use only under 10 million USD of revenue.
A softer cap than Stability's, and a real one. It is the kind of clause that costs nothing on day one and forces a migration on the day you succeed.
What we did instead: Usable, but do not build your only pipeline on it.
check it yourself ↗ · read by us on 2026-09-18
Hyper-SD
Three different licences glued into one file.
Which one applies depends on which checkpoint you downloaded. The FLUX variants are non-commercial; others are not. One repository, one licence file, several answers.
What we did instead: Find the line for your exact checkpoint before you ship anything.
check it yourself ↗ · read by us on 2026-09-18
LCM_Dreamshaper
The label says MIT, but it is distilled from SD1.5.
The metadata tag on the page and the provenance of the weights disagree. Distillation carries the original terms with it — here, OpenRAIL — so the tag is the least reliable thing on that page.
What we did instead: Read where the weights came from, not the tag.
check it yourself ↗ · read by us on 2026-09-18
Sana 0.6B
Not a licence trap - a SIZE trap. The 0.6B pulls a closed 5.2 GB encoder.
The number in the name describes one part. The package you actually have to ship is about 16.5 GB, and the biggest piece of it is not open. If you picked it because it sounded small, you picked it on a number that was true and irrelevant.
What we did instead: Weigh the whole chain, not the headline parameter count.
check it yourself ↗ · read by us on 2026-09-18
PixArt-alpha
Same shape: a 9.5 GB T5-XXL encoder rides along.
Real total lands between 20 and 31 GB depending on what you keep. The model itself is not the thing that decides whether it fits your card.
What we did instead: Budget for the encoder before you budget for the model.
check it yourself ↗ · read by us on 2026-09-18
Questions people actually ask
Can I use Qwen2.5-VL-3B commercially?
No. The 3B carries a non-commercial licence while the 7B and 72B in the same family are Apache-2.0. The smallest model, the one that fits a modest card, is the one you cannot build a paid product on. Read on 2026-09-18.
Is SDXL-Turbo free for commercial use?
Only below one million USD of revenue, and only if you display 'Powered by Stability AI' on the product itself. SD-Turbo carries the same terms. Read on 2026-09-18.
Why is my 0.6B model a 16 GB download?
Because the parameter count in the name describes one part of the chain. Sana 0.6B pulls a closed 5.2 GB text encoder; PixArt-alpha pulls a 9.5 GB T5-XXL. The headline number is true and irrelevant - weigh the whole package.
Does a distilled or fine-tuned model inherit the original licence?
In practice, yes, and that is where two of these traps live. Nanonets-OCR-s descends from a non-commercial model, and LCM_Dreamshaper is tagged MIT while being distilled from SD1.5. Read where the weights came from, not the tag on the page.
How did you check these?
We read the licence file or the model card for each one and wrote down what it says, with the date. We did not run most of these models, and we say so on their cards here - they carry the REVIEWED label rather than a measurement. This is not legal advice; every row links to the source so you can read it yourself.
Why publish what you rejected?
Because everyone publishes what they chose. The reason we did NOT take something is the part that would have saved us a week, and it is not written down anywhere else.
Why publish what we rejected? Everyone publishes what they chose. The reason we did not take something is the part that would have saved us a week — and it is not written down anywhere else. The same thinking runs through the shelf, where every tool says whether we measured it, ran it, or only weighed it on paper.