Use Qwen3.8-27b?

[kripper]

🚀 The feature

Have you considered using/supporting Qwen3.8-27b?

Motivation, pitch

I got good results with Qwen3.8-27b to extract texts

Alternatives

No response

Additional context

No response

[felixdittrich92]

Hi @kripper 👋,

We currently don't have any VLM support planned.

The main reason is that, while models like Qwen can produce very good text extraction results, they generally don't provide the same level of structured OCR information that we need. For example, we rely on things like word/line coordinates, confidence scores, document layout, reading order, and other intermediate information to make the results useful for downstream document processing.

With a VLM, you typically get a textual answer from the model, but you don't necessarily get reliable or deterministic information about where a particular piece of text came from or how confident the model is in that prediction. This essentially means that the user has to trust the model's output without having the same level of transparency and control.

For a general-purpose text extraction use case, I definitely see the appeal of Qwen3.8-27B, and it's great to hear that you're getting good results with it. However, for docTR our focus is currently on models that provide these structured and more explainable outputs.

That said, if there is a reasonable way to combine VLMs with reliable spatial information, confidence estimates, and the other metadata we expose, I think it would be an interesting direction to explore in the future. 🚀