PickXiaomi-OCR-0: OmniDocBench 92.7 points, 0.8B document OCR

A vision-language model specialized in extracting text and understanding structure from document images. Fine-tuned based on Qwen3.5-0.8B-Base, it performs complex document understanding tasks despite being a small model. It uses a multimodal architecture that processes images and text together.

1 / 7

It achieved a composite score of 92.73 on PaddlePaddle's Real5-OmniDocBench benchmark. It demonstrates consistent performance of over 91 points even under various real-world document conditions such as scanning, distortion, screen capture, lighting, and tilt. It focuses on understanding document structure beyond simple Optical Character Recognition.

Key features include document parsing, document understanding, visual question answering, and key information extraction. It supports an interactive interface, allowing users to ask questions by mixing images and text. It is compatible with the Transformers library and is released as open source.

Thanks to its lightweight model structure, it is suitable for building document processing pipelines in local environments or on edge devices. It is useful for developers who need high-accuracy document information extraction without using large-scale LLMs.