PP-OCRv6 on Hugging Face: Text Recognition in 50 Languages with Models from 1.5M to 34.5M Parameters

The PaddleOCR team has introduced PP-OCRv6 — a new generation of universal OCR models for text detection and recognition in real-world scenarios: documents, screenshots, multilingual images, digital displays, industrial labels, and street text. The family includes three tiers — tiny (1.5M parameters), small (7.7M), and medium (34.5M), with medium and small supporting 50 languages, including simplified and traditional Chinese, English, Japanese, and 46 Latin-based languages. On internal PaddleOCR benchmarks, the PP-OCRv6_medium model achieves 86.2% Hmean for detection and 83.2% recognition accuracy, which is 4.6 and 5.1 percentage points higher than PP-OCRv5_server, respectively. For detection, RepLKFPN is used — a lightweight feature pyramid with large convolution kernels — while recognition employs EncoderWithLightSVTR, combining local context modeling with global attention. Developers have access to several inference backends: Paddle Inference, Transformers (via Hugging Face), and ONNX Runtime. Python code examples show how to run OCR with different backends. Models are published on the Hugging Face Hub in safetensors, Paddle, and ONNX formats, and an online demo is also available.