PaddleOCR Releases PP-OCRv6 OCR Model Family Supporting 50 Languages
멀티모달 | Tue Jun 23 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 1 sources
PaddleOCR released PP-OCRv6, a lightweight OCR model family ranging from 1.5M to 34.5M parameters, on Hugging Face.
Analysis
[PaddleOCR] released PP-OCRv6 OCR model family on Hugging Face [1]
- Available in three tiers: tiny (1.5M)
- small (7.7M)
- and medium (34.5M)
- Supports 50 languages including Simplified/Traditional Chinese
- English
- Japanese
- and 46 Latin-based languages
- Handles real-world scenarios such as documents
- screenshots
- digital displays
- and industrial labels
- Compatible with PaddlePaddle
- Transformers
- and ONNX Runtime backend integrations
[PP-OCRv6_medium] delivered significant improvements in detection and recognition accuracy over the previous generation [1]
- Achieved 86.2% Hmean in text detection
- Achieved 83.2% accuracy in text recognition
- Improved detection by +4.6%p and recognition by +5.1%p over PP-OCRv5_server
- Suitable for server-side pipelines and industrial OCR
[PPLCNetV4 and RepLKFPN] introduced unified backbone and large-kernel FPN architecture [1]
- Adopted PPLCNetV4 as a unified backbone for both detection and recognition
- Enhanced multi-scale text detection with RepLKFPN
- Improved handling of small
- rotated
- or low-resolution text
- All tiers from tiny to medium share the same architectural direction
[EncoderWithLightSVTR] applied a lightweight text recognition module [1]
- Uses a LightSVTR-based encoder in the text recognition stage
- Emphasizes the practicality of specialized OCR models even in the VLM era
- Optimized for generating structured text output
- Supports flexible deployment across mobile
- desktop
- and edge devices