Qwen-Image-3.0 has no open weights, so we tested the ones that do

What we checked

Alibaba announced Qwen-Image-3.0 on 2026-07-21. The pitch was a 4,500-token prompt window, twelve languages, and text that stays readable down to roughly ten pixels, which is the difference between a model that decorates a layout and one that can set it. The release carried no benchmark table, no model card and no technical report, and access was invite-only through the API.

That leaves one question you can settle with a query instead of an opinion: is there anything to download? We asked both hubs on 2026-08-13, then listed the two organisations in full so a wrong guess at a repo name could not fake the answer.

The open Qwen line did not stop

Nothing was withdrawn. Qwen-Image has been Apache-2.0 since 2025-08-02 and still moved about 190,000 downloads in the last 30 days. The December point releases sit beside it under the same licence, and Qwen-Image-Layered is there too.

So the honest version of this story is narrower than "Alibaba went closed". Dated point releases keep arriving with weights attached. The flagship line is what stopped, and it stopped for images specifically. The Qwen3.8-2.4T-A95B repo went up on 2026-08-08 and was still filling in on 2026-08-12; when we counted on 2026-08-13 it held 213 weight shards under a custom Qwen licence. That week the text flagship opened and the image flagship did not.

Where the open models actually break

Whether a 6B open model matches a hosted flagship is not the interesting question, because it does not. The useful question is which jobs go from impossible to routine when you pay, since those are the ones worth paying for.

Dense text is the clearest of them. Frame 01 asked for a newspaper front page and produced a headline reading FRONT IEIM, then three columns with the greyness of body copy and no words in them. Prompt work does not move this.

Frame 02 is more instructive because it half worked. The grid came out correct, nine cells with nine captions in the right places, and five of those captions spell Coffee. The other four spell nothing. A layout the model can build but cannot label is exactly the gap the hosted flagships claim to have closed.

Frame 05 shows a third failure that is easy to miss. Every label is spelled right, 33 letters of it, and the diagram is still wrong: evaporation points down into the sea, precipitation points up at the mountain. Spelling a word is not the same as knowing what it means.

The boundary is size, not language

Two frames came back clean, and together they locate the line.

Frame 03 put twenty characters on a jam jar across two lines and got all twenty right. Frame 04 is the one that surprised us: the two Chinese characters on the shop board are correct and well proportioned, while the small red placards a few feet behind them in the same image are stroke-shaped noise. One of us reads Chinese, which is the only reason we tested a script we can check. We did not test Korean or Arabic, because we could not have judged the result.

One image, one model, one run, both outcomes. Everything we asked for at sign size came back right. Everything at caption size or below came back wrong. That, rather than any language boundary, is what separates the work these models can do from the work they cannot.

If you need the dense-text work today

Nothing free and open does it as of 2026-08. If the job is a newspaper page, a dense infographic, or a poster whose small print has to survive a print shop, the real options are hosted and paid: Qwen-Image-3.0 through Alibaba's API, GPT Image 2, or Nano Banana Pro. We are the wrong tool for that job and no prompt makes us the right one.

The workaround is free and covers most of what people actually need. Generate the picture with no text in it, then set the type yourself in a real font. Frame 06 is that test: one prompt, a usable image, and an empty band across the top waiting for a headline.

Our ceiling belongs in the same paragraph. Output here is 1024x1024, 1152x864, 864x1152, 1280x720 or 720x1280, which sits under a modern phone screen at native size. Anything going to print or a large display needs upscaling afterwards.