On May 20, 2025, the PaddleOCR team from Baidu officially released version 3.0 of PaddleOCR and made it open source. This new version has made significant progress in text recognition accuracy, multilingual support, handwriting recognition, and high-precision document parsing, further enhancing PaddleOCR's technical strength and application value in the OCR field.
Since its release, PaddleOCR has been widely loved by academia, industry, and research due to its cutting-edge algorithms and industrial implementation practices, and has been extensively applied in numerous well-known open-source projects. The newly released PaddleOCR 3.0 fully supports the official version 3.0 of the PaddlePaddle framework. It not only improves text recognition accuracy but also supports multiple types of text recognition and handwriting recognition, meeting the demand for complex document high-precision parsing in large model applications. In addition, combined with the Wenxin Qianwen 4.5 Turbo large model, PaddleOCR 3.0 significantly enhances key information extraction accuracy and adds support for domestic hardware such as Kunlun Chip and Ascend.
One of the core highlights of PaddleOCR 3.0 is the full-scene text recognition model PP-OCRv5. This model supports five types of text: simplified Chinese, traditional Chinese, Chinese pinyin, English, Japanese, as well as complex text scenarios such as handwriting, vertical text, pinyin, and rare characters. Compared to the previous generation, PP-OCRv5 has improved overall recognition accuracy by 13 percentage points, reaching a leading level in the industry. PP-OCRv5 achieves seamless recognition of multiple text types through a unified model architecture, simplifying the deployment process and improving overall recognition accuracy and speed.
In terms of document parsing, PaddleOCR 3.0 introduced the general document parsing solution PP-StructureV3. This solution strengthens the capabilities of layout area detection, table recognition, and formula recognition, adding chart understanding and multi-column reading order recovery capabilities, and can convert results into Markdown and JSON formats. PP-StructureV3 leads many open-source and closed-source solutions in the OmniDocBench benchmark test, showcasing its powerful ability in high-precision parsing of multi-scenario, multi-layout PDFs.
In addition, PaddleOCR 3.0 also released the intelligent document understanding solution PP-ChatOCRv4. This solution natively supports Wenxin Qianwen 4.5 Turbo, and the key information extraction accuracy has improved by 15 percentage points compared to the previous generation. PP-ChatOCRv4 combines the advantages of large models and small models, supporting offline use of the multimodal document understanding model PP-DocBee2, enabling a one-stop solution for complex document information extraction problems such as layout analysis, rare characters, multi-page PDFs, tables, and seal recognition.
The release of PaddleOCR 3.0 not only demonstrates Baidu's continuous innovation in OCR technology but also provides developers with more powerful and user-friendly tools, helping them quickly implement AI applications.
Open Source Address:
https://github.com/PaddlePaddle/PaddleOCR