Descriptions:
Fahd Mirza tests GutenOCR-3B, a vision-language model fine-tuned from Qwen 2.5V specifically for document OCR tasks. Unlike traditional OCR tools that return plain text, GutenOCR produces ‘grounded OCR’ — it returns each word or line alongside its precise bounding box coordinates, enabling targeted queries such as reading only the text within a specific region of a page. The model is available in 3B and 7B parameter versions on Hugging Face.
Mirza installs the model via the UV package manager on Ubuntu with an Nvidia RTX 6000 (48 GB VRAM), downloading a roughly 7 GB model file on first run. Testing reveals the model performs adequately on simple English and Chinese text but fails significantly on multicolumn layouts, tables, invoices, and mathematical formulas. He also identifies two architectural shortcomings noted in the project’s paper: the fine-tuning process caused catastrophic forgetting of base model capabilities, and the model prioritizes layout-preserving reading order over canonical markdown conversion, which can inflate character error rates even when all content is captured.
Mirza’s conclusion is to stick with Qwen 2.5 VL or the newer Qwen 3 VL for diverse document processing needs, while flagging GutenOCR as a project worth watching for future improvements in its targeted use cases.
📺 Source: Fahd Mirza · Published February 21, 2026
🏷️ Format: Review
