MonkeyOCRv2 is a document-oriented visual-text foundation model designed to preserve dense text, character strokes, and layout details that natural-image encoders often miss. Its MonkeyDoc v2 corpus contains 113 million document images across 17 languages. The model jointly trains image-to-text generation with pixel-level document reconstruction. Replacing existing encoders improves five tasks: text recognition, formula recognition, text detection, tampering detection, and overlapping-text segmentation. A frozen encoder paired with a lightweight language model produces a 0.7B document parser that reportedly surpasses the previous 3B dots.mocr system on MDPBench by 2.8 absolute points, while also improving document understanding against CLIP-, DINO-, and SAM-based counterparts.
No heat snapshots are available in the last 24 hours.