[Badcase]: Qwen-Chat 3: Polish grammar errors & garbled Polish text on images

[gileneusz]

Model Series

Qwen3

What are the models used?

Qwen3-235B-A22B-2507

What is the scenario where the problem happened?

Qwen Chat on https://chat.qwen.ai/

Is this badcase known and can it be solved using avaiable techniques?

• I have followed the GitHub README.
• I have checked the Qwen documentation and cannot find a solution there.
• I have checked the documentation of the related framework and cannot find useful information.
• I have searched the issues and there is not a similar one.

Information about environment

Qwen Chat on https://chat.qwen.ai/

Description

First of all—great work: Qwen-Chat 3 shows a clear improvement over Qwen 2 / 2.5 for Polish (reasoning, instruction following, overall usefulness). Two issues remain in chat, though: (1) Polish still has grammatical/inflection errors (which I don’t see with OpenAI or Google Gemini on the same prompts), and (2) the image generator understands Polish prompts, but fails to render correct Polish text on images; additionally, when a prompt references “the story above,” prior chat context often isn’t passed into the image tool.

Steps to reproduce

This happens to Qwen3-235B-A22B-2507.
The bad case can be reproduced with the following steps:

1. In chat, request a short story in Polish.

2. Observe that the Polish output occasionally contains morphology/inflection mistakes (case endings, agreement).

3. Then ask the model to generate an image with Polish dialogue taken from the story above (“…z dialogami z powyższej historii”).

4. Observe two problems:

   • Context hand-off: the image generator does not use the prior chat story unless you paste it inline again.
   • On-image text: even when it draws the right scene, the Polish text on the image is garbled (broken diacritics, pseudo-Slavic gibberish).

The following example input & output can be used:

system: You are a helpful assistant.

user: Napisz mi historię o mrówce i słoniu. (write a story about an ant and an elephant)

assistant (example, Polish story): "Mrówka, która chciała być słoniem" … [long story in Polish language]

user: Wygeneruj obraz z dialogami z powyższej historii. (generate an image with dialogs from the story above)

assistant (image result): The picture is not about the story above; Polish text is wrong
(ie: wrong spelling, incomprehensible words).

Expected results

The results are expected to be:

• Grammatically correct, idiomatic Polish in chat (proper case endings, agreement, diacritics).
• Image generation that (a) correctly uses prior chat context when referenced, and (b) renders valid Polish text on images (real words with correct spelling and diacritics).

Attempts to fix

Attempts to fix

I tried simply specifying in Polish exactly what text should appear on the image. For example, I wrote:

napisz w tym rysunku o tej historii: (write in the picture about the story)

- hej jak się czujesz?
- czuję się dobrze

but the generated on-image text was (totally not Polish language - not even any language):

Hej, jak se, yei cckuzzes?
Mh
Czújo s
dokbrce

And the picture is not about the ant and the elephant

Anything else helpful for investigation

• Compared to OpenAI and Google Gemini, I don’t encounter these Polish morphology issues on equivalent prompts.
• Despite the above, Qwen-Chat 3 is a big step forward vs 2 / 2.5. My guess is that high-quality Polish data is still relatively limited; please consider increasing Polish training data (especially morphology-rich text) and strengthening chat→image context hand-off and Polish text rendering in the next iteration.