Baidu releases open OCR model for reading long documents at once
Baidu released an called Unlimited-OCR. It is meant to read long PDFs or multi-page documents without cutting them into separate pages and stitching the text back together later. The main idea is to keep the original document image available as a reference while making the model remember only a small moving part of the text it has already produced.
This can reduce the growth of the KV cache, which otherwise can fill VRAM during long document runs. Lower memory use may make it possible to run on the same machine or process long documents more reliably. The release is mainly about OCR and , not a complete answer for analyzing or summarizing very large documents.
The includes examples for , an SGLang server, PDF-to-image conversion, and with infer.py. The tested setup uses an NVIDIA GPU and a CUDA-based Python environment.
Key points
- Baidu released Unlimited-OCR as an document reading model.
- It targets one-shot parsing of long PDFs and multi-page documents.
- Its design aims to reduce KV cache growth and lower VRAM pressure.
- The repo includes , SGLang, PDF, and examples.
- The release is mainly useful for OCR, not a full long-document reasoning system.