PP‑OCRv6: 50‑Language OCR from 1.5M to 34.5M Parameters
Deploy PP‑OCRv6 medium tier for multilingual OCR; it offers 86.2 % detection and 83.2 % recognition, and supports Paddle, Transformers, and ONNX backends.
Download the PP‑OCRv6 medium weights from Hugging Face, load them with PaddleOCR, and test on your multilingual documents.
Summary
On June 22 2026, PaddleOCR released PP‑OCRv6, a universal OCR model family scaling from 1.5 M to 34.5 M parameters across tiny, small, and medium tiers. The medium tier supports 50 languages, including Simplified Chinese, Traditional Chinese, English, Japanese, and 46 Latin‑script languages, achieving 86.2 % detection Hmean and 83.2 % recognition accuracy on PaddleOCR’s official benchmarks. Compared to PP‑OCRv5_server, PP‑OCRv6 improves detection by +4.6 pp and recognition by +5.1 pp. The architecture uses a PPLCNetV4 backbone, RepLKFPN for lightweight multi‑scale detection, and EncoderWithLightSVTR for recognition, combining local context modeling with global attention. The model family is available in multiple formats: safetensors, Paddle inference, and ONNX, and can be run via PaddleOCR’s Transformers, ONNX Runtime, or native Paddle inference backends. The online demo and Hugging Face Space allow quick evaluation, while the collection provides ready‑to‑use weights for developers. PP‑OCRv6’s lightweight tiers make it suitable for edge devices, mobile, and server‑side OCR pipelines. The release also includes a structured JSON output for downstream document parsing, search, and RAG workflows. The new OCR family positions PaddleOCR as a competitive choice for multilingual text extraction in real‑world applications.
Key changes
- Release of PP‑OCRv6 with tiny (1.5 M), small (7.7 M), and medium (34.5 M) tiers.
- Medium tier supports 50 languages and achieves 86.2 % detection Hmean, 83.2 % recognition accuracy.
- Detection improvement over PP‑OCRv5_server: +4.6 pp; recognition improvement: +5.1 pp.
- Architecture uses PPLCNetV4 backbone, RepLKFPN detection, EncoderWithLightSVTR recognition.
- Supports Paddle, Transformers, and ONNX Runtime backends; weights in safetensors, Paddle, ONNX.
- Provides structured JSON output for downstream parsing, search, and RAG.
- Online demo and Hugging Face Space for quick evaluation.
- Lightweight tiers suitable for edge, mobile, and server‑side OCR pipelines.