Redaction with Qwen 3.5
Overview
In February 2026, I wrote an article looking into the use of VLMs for OCR and redaction tasks in documents. The original article, using Qwen 3 VL 8B Instruct, can be found here. In late February 2026, Qwen 3.5 was released and I have been testing out the same examples discussed in the original blog post. Here I will show a comparison between Qwen 3 and Qwen 3.5 for OCR/redaction tasks with an updated version of the app (version 2.0.1, which you can find here).
My conclusion from the last article was that PaddleOCR for initial OCR, paired with Qwen 3 VL 8B Instruct for low confidence phrases was the best solution for ‘difficult’ pages in documents. How does this conclusion hold up with Qwen 3.5? Let’s take a look.
The Document Redaction VLM space on Hugging Face can be found here. By default, Qwen 3.5 27B is loaded in as the VLM model, quantised to 4 bit using bitsandbytes.
Example 1: Difficult handwritten note
The first example is a note with handwriting that is hard to decipher even for a person.
As a reminder, Qwen 3 VL 8B Instruct alone found this:
A hybrid approach of PaddleOCR + Qwen 3 VL 8B Instruct found this:
Now let’s try Qwen 3.5 27B.