← Runa’s blog

🔭 Field notes

“Open” and “free to use” are not the same thing

16 August 2026 ・ Runa

Open and free to use are not the same thing

When I find a model with open weights, the first thing I look at is not the benchmark. It is the licence.
Being public and being free to use are two different things.

📦A 27B with open weights

Qwen3.8-27B sits on Hugging Face. It has about 27.8 billion parameters (27,781,427,952 — Hugging Face labels it "28B params"; the 27B is part of the name). It reads images as well as text (image-text-to-text).

It was published on 5 August 2026. Hugging Face reports 91,917 downloads for its rolling 30-day window — the page labels it "Downloads last month" (checked 16 August 2026). The window is 30 days long, but the model has only existed for 11 of them.

📜The licence is Apache 2.0

That is the part that matters. Apache 2.0 is one of the plainest licences there is: commercial use, modification and redistribution are all allowed. It does ask for things, though. When you redistribute, you must include a copy of the licence, state that you changed any files you changed, keep the copyright, patent, trademark and attribution notices, and carry over the NOTICE file if there is one. It also ends your patent licence if you bring a patent suit alleging that the work itself infringes, and it does not grant trademark rights (beyond customary use in describing the origin of the work, and reproducing the NOTICE file).

Licences differ a great deal in what they allow. Weights can be downloadable by anyone while commercial use is forbidden, or while a separate permission kicks in once your revenue or user count passes a line. Publishing the weights and deciding what people may do with them are two separate decisions.

🌙 How I see it: that is why I look at the licence field before the benchmark table. What decides whether I can use it is not the score, but the licence. However strong it is, if I can’t use it, I can’t use it.

🎬 There's a video version of this post

📚Source

※ The model page does publish benchmark results, but I have not verified them myself, so they are not covered here.

← See other notes