MonkeyOCRv2 with 0.7B Parameters Tops Open-Source Document OCR, Surpassing 3B Models

A year ago, 3B parameters could still be called 'lightweight document models.' Now, MonkeyOCRv2-S (0.6B) and MonkeyOCRv2-B (0.7B) score 82.5 and 83.3 on MDPBench, respectively, surpassing the previous top open-source model dots.mocr (3B, 80.5). The two smallest models on the leaderboard now stand at the forefront of open-source results.

MonkeyOCRv2-B consists of a 0.1B visual encoder and a 0.6B language model, while dots.mocr has approximately a 1.2B visual encoder plus a 1.8B language model. The former has less than a quarter of the total parameters, and the visual encoder is only about one-eleventh the size. The team did not simply prune; they reallocated responsibilities: the visual side provides clean and complete page evidence, and the language side no longer has to do extensive character completion or guessing.

The 0.7B version achieves 81.7 on photographed documents and 82.1 on non-Latin scripts, indicating that the advantage is not limited to well-formatted PDFs or common Latin scripts.

The fundamental difference between documents and natural images is that details determine meaning: '王' and '玉' differ by only a dot, and 'o' and '0' have different aspect ratios. In a case study from the paper, an infographic actually reads '700,000,' but various general visual encoders read it as '100,000.' Therefore, MonkeyOCRv2 adopts two complementary pretraining objectives: image-to-text generation learns 'what the page says,' and pixel-level document reconstruction requires visual tokens to retain enough information to recover character strokes, formula symbols, and layout structure.

Reconstruction is not the final output of a reconstructed image; rather, the reconstruction training enables a smaller visual encoder to lose fewer critical details when compressing a page. The team built MonkeyDoc v2, containing 113 million document images covering 17 languages, including 8 million full pages and 105 million fine-grained elements, with about 61 million real samples and 52 million synthetic samples. The data has been open-sourced, allowing the community to retrain from the same starting point.

The 17 languages correspond to different fonts, reading directions, and real document forms, not multilingual translations of the same template. The model learns not just more characters, but more ways in which 'a page is constructed.'

The MonkeyOCRv2 visual encoder was plugged into seven tasks: text recognition, formula recognition, text detection, document tampering detection, overlapping text segmentation, document parsing, and document understanding, yielding consistent improvements. In traditional text recognition, the CRNN composite result improved from 58.7 to 67.3; on common OCR benchmarks, the average improvement was 2.3 percentage points.

In formula recognition, after replacing with the MonkeyOCRv2-S encoder, the approximately 110M-parameter UniMERNet-T surpassed the 325M-parameter UniMERNet-B in overall exact match rate, achieving 'small model surpassing large model.' Text detection, document tampering detection, and overlapping text segmentation improved by 3.3, 7.5, and 5.3 percentage points, respectively.

A single backbone that simultaneously improves recognition, detection, segmentation, parsing, and understanding indicates that it has learned not a decoder-specific trick, but a set of document visual representations that can be reused by different systems.

DeepSeek-OCR proposes optical context compression, advancing the efficiency issue from 'model parameters' to 'how many visual tokens a page needs.' Its paper reports that using only 100 visual tokens can surpass GOT-OCR2.0; with fewer than 800 tokens, it surpasses MinerU2.0, which uses an average of over 6,000 tokens. But what the compressed tokens retain is key.

MonkeyOCRv2's 'reconstruction as visual memory' directly answers this question: if visual tokens can still recover character strokes, glyph outlines, and layout structure, it means details have not been easily forgotten. The number of visual tokens measures the degree of compression; reconstruction capability tests how much reliable evidence remains after compression.

From MonkeyOCR's 3B to MonkeyOCRv2's 0.7B, the path is not simple weight loss, but a continuous redefinition of module responsibilities. The final delivery is not just higher scores, but a reproducible, comparable, and improvable set of document parsing capabilities.