Qwen-Image-2.1 shipped under a licence that grants use for research or evaluation only, which is why a lot of people wrote it off for paid work. We covered that licence in September, noted that it names an address to ask at, and said asking was a real option rather than a polite fiction.
This morning somebody posted what happened when they asked. They emailed the licensing address in September and got back a tier structure, by images generated per calendar month:
| images per calendar month | reported cost |
|---|---|
| up to 300,000 | free |
| 300,000 to 3 million | a 20 percent revenue share |
| above 3 million | 500,000 US dollars a year plus 15 percent |
Their own framing is the useful one. They generate 20 to 30 images a month. A shop producing 100 a day lands at about one percent of the free tier. Unless you are running a generation service, you are inside the first line.
What this is
One person’s account of private correspondence, posted in good faith, and unverifiable by us or by you.
We went looking for those numbers published anywhere and did not find them. Not on a pricing page, not in the model card, not in the repository. The poster says the same thing, that none of it is documented publicly, which is the honest part of their post and the reason to read it carefully rather than dismiss it.
The licence file itself has not changed. It still says the grant is for non-commercial purposes, still defines that as research or evaluation, and still points at an email address for anything else.
So an inbox says one thing and the document says another, and only one of those is a document.
Why that distinction is doing real work
An email from a licensing desk is not an amendment to a licence. It may well reflect exactly how the company intends to behave, and the person who received it is probably safe in practice. But if you are earning money from something, what protects you is a written grant that names you, not a screenshot of somebody else’s reply.
That is not a reason to ignore the post. It is a reason to treat it as what it is: strong evidence that asking produces a generous answer, and a prompt to go and ask for yourself if the answer matters to your work. The address is in the licence. Based on this account, the reply takes a few weeks and is friendlier than the text implies.
Meanwhile there is a whole layer of secondary coverage with titles along the lines of open weights you cannot ship, and licence bars commercial use. Those pieces are reading the document correctly. They are also now, if this report holds, describing a situation that does not match how the licensor actually behaves, and no reader can tell the difference because the generous version is undocumented.
That is the fourth time in three weeks that the distance between a headline and the underlying fact has decided what people believe. The other three ran in the direction of alarm. This one runs the other way, and it is harder to correct, because the generous version exists only in one person’s inbox.
Where we stand
We use Qwen-Image-2.1 output in what we make, we are several orders of magnitude below the first tier, and our position on it was settled before any of this surfaced. Nothing here changes what we do.
What we would say to anyone deciding: the numbers above are reported, not confirmed, and we have not independently verified them. If your work depends on the answer, send your own email rather than trusting ours or anybody else’s.
A daily note from our news radar. The week gets the full treatment in the DIY AI Brief every Monday.