PP-OCRv6 integrates multiple languages into a single model, covering everything from edge devices to the cloud. Domestic OCR is finally gaining strength.

Baidu PaddlePaddle launches PP-OCRv6. The new version offers three model options: tiny 1.5M, small 7.7M, and medium 34.5M, covering edge, browser, and cloud scenarios. Compared with v5, detection and recognition accuracy improve by 4.6% and 5.1%, respectively, and Chinese, English, Japanese, and 46 Latin languages are aggregated into a single model. The newly designed detection/recognition network introduces a unified module and structural re-parameterization to enhance accuracy and reduce computational requirements. With OpenVINO optimizations, end-to-end CPU inference on the medium model sees up to a 5.2x improvement, and the code has been integrated into PaddleOCR and open-sourced.