PaddleOCR выпустила PP-OCRv6: распознавание текста на 50 языках от 1,5 до 34,5 млн параметров

Baidu PaddlePaddle has released PP-OCRv6 on Hugging Face — a new generation of universal OCR models that recognize text in 50 languages within a single architecture. Performance improvements in key metrics compared to the previous server version range from 4.6 to 5.1 percentage points, and the lower boundary of the model family fits within 1.5 million parameters.

Three Configurations for Different Tasks

PP-OCRv6 comes in three variants: Tiny, Small, and Medium. The parameter range — from 1.5 to 34.5 million — covers the spectrum from embedded systems with strict memory constraints to server pipelines where maximum accuracy is critical.

Key metrics for the three configurations:

• Tiny (1.5 million parameters) — detection 80.6% Hmean, recognition 73.5%
• Small (7.7 million parameters) — detection 84.1% Hmean, recognition 81.3%
• Medium (34.5 million parameters) — detection 86.2% Hmean, recognition 83.2%

The Small and Medium versions support 50 languages within a single model: simplified and traditional Chinese, English, Japanese, and 46 Latin-based languages. This eliminates the need to maintain and update separate language models — one of the major operational complexities in production OCR pipelines serving global audiences.

The Tiny version is designed for scenarios where inference speed is the priority under limited computational resources, and comprehensive language coverage is not critical.

Architecture: Large Kernels and Lightweight Attention

All three configurations are built on a unified backbone PPLCNetV4 that combines text detection and recognition tasks. Unification reduces maintenance costs and simplifies transitions between model sizes without rebuilding the pipeline.

Text detection uses RepLKFPN — a lightweight feature pyramid based on large convolutional kernels. This design enables simultaneous handling of text at different scales: from small fonts in legal documents to large characters on industrial labels and street signs.

Recognition is handled by EncoderWithLightSVTR — a hybrid of local contextual modeling and global attention mechanisms.

The models were tested on a wide range of industrial scenarios: business documents, interface screenshots, price tags, digital displays, signs, and text in natural scenes. Compared to PP-OCRv5_server, improvements were +4.6 percentage points in detection and +5.1 percentage points in recognition.

Three Paths to Production

PaddleOCR 3.7 provides a unified API for three deployment backends:

• Transformers — native integration with Hugging Face Hub and PyTorch pipelines without additional configuration
• ONNX Runtime — cross-platform format with no framework dependencies; convenient for heterogeneous infrastructure mixing Python, C++, and mobile clients
• Paddle Inference — native format for maximum performance in Baidu infrastructure

A collection of 19 models has been published on Hugging Face: safetensors versions, Paddle inference files, and ONNX variants — separate checkpoints for detection and recognition at each size.

Inference returns two types of data: structured JSON with bounding box coordinates and recognized text, as well as an image with visualization for tasks requiring visual verification.

For quick testing without package installation, an interactive demo is available on Hugging Face Spaces.

What This Means

PP-OCRv6 addresses two practical challenges simultaneously: multilingual coverage without proliferation of separate models and deployment flexibility without tight coupling to the PaddlePaddle ecosystem.

Availability on Hugging Face makes the library accessible to any Python team — for multi-language document tasks, it is now one of the most compact and well-documented options in open source.