Huazhong University · 2026-07-13 · notable
MonkeyOCRv2 — visual-text foundation model for document AI
MonkeyOCRv2 is a visual-text pretrained model for document parsing, released in three sizes (S/B/AS) with open weights, inference code, and a demo. Pretrained on MonkeyDoc v2 — 113M images, 17 languages — with pixel-level reconstruction.
MonkeyOCRv2 pretrains a visual-text foundation model on 113M multilingual document images.
What is it?
MonkeyOCRv2 is a visual-text pretrained model for document AI, released by Yuliang Liu and Xiang Bai's group at Huazhong University. Three variants ship with open weights: MonkeyOCRv2-S, MonkeyOCRv2-B, and MonkeyOCRv2-AS, alongside inference scripts and a web demo.
How does it work?
Training pairs image-to-text generation with pixel-level reconstruction so the encoder preserves fine-grained character detail. Pretraining data is MonkeyDoc v2, described as the largest document-image corpus so far — 113M images across 17 languages. The vision encoder plugs into multimodal LLMs as a frozen backbone.
Why does it matter?
MonkeyOCRv2 beats the 3B-parameter dots.mocr baseline by 2.8% on MDPBench with an ~11× smaller vision encoder, so document-AI systems can shrink without losing accuracy. Open weights plus code cover the five main document tasks — parsing, layout, structure, table, and formula — which is uncommon in one release.
Who is it for?
Document-AI researchers and teams building OCR / RAG pipelines over multilingual PDFs.
Try it
git clone https://github.com/Yuliang-Liu/MonkeyOCRv2