500M Params Match Billion-Level Models: Baidu PaddleOCR Tops GitHub OCR Projects

Baidu PaddlePaddle's open-source OCR library PaddleOCR has reached 73,300 GitHub stars, surpassing Google's Tesseract (73,200 stars) to become the top OCR project on GitHub. MinerU ranks third with 57,500 stars. Since its open-source release in 2020, PaddleOCR has supported over 100 languages and covered more than 160 countries and regions.

Recent updates to PaddleOCR are frequent: The newly released PP-OCRv5 last week, with only 500 million parameters, matches the accuracy of billion-level vision-language large models on standard OCR benchmarks. PaddleOCR-VL-1.5 also set a new record with 94.5% accuracy on the document parsing benchmark OmniDocBench v1.5.