PaddleOCR:基于飞桨生态的OCR与文档智能引擎项目

PP-OCRv6 highlights:

• Accuracy boost: Medium tier achieves +4.6% detection and +5.1% recognition over PP-OCRv5_server, surpassing mainstream VLMs (Qwen3-VL-235B, GPT-5.5) with only 34.5M parameters.
• 50 languages unified: Single model covers Chinese, English, Japanese, and 46 Latin-script languages — no model switching needed.
• Specialized scenarios: Major improvements in digital displays, dot-matrix characters, tire prints, and industrial text recognition.
• Faster inference: 5.2× CPU speedup (OpenVINO), 6.1× on Apple M4 (tiny), 0.13s on A100 GPU.
• Three tiers for all scenarios: tiny (1.5M) / small (7.7M) / medium (34.5M) for edge, mobile, and server deployment.
• Model availability: All models are available on HuggingFace and ModelScope.

PaddleOCR-VL-1.6 highlights:

• New SOTA Accuracy: Achieves over 96.3% on OmniDocBench v1.6, also sets new SOTA on OmniDocBench v1.5 and Real5-OmniDocBench, leading both open-source and proprietary solutions in text, formula, and table recognition.
• Comprehensive Capability Upgrade: Significant improvements in table, ancient document, and rare character recognition, with notably enhanced seal recognition, spotting, and chart understanding across multiple scenarios.
• Seamless Migration: Model architecture is fully consistent with PaddleOCR-VL-1.5, enabling zero-cost adaptation—swap and go.
• Try it now: Available on HuggingFace or our Official Website.

• 94.5% Accuracy on OmniDocBench: Surpassing top-tier general large models and specialized document parsers.
• Real-World Robustness: First to introduce the PP-DocLayoutV3 algorithm for irregular shape positioning, mastering 5 tough scenarios: Skew, Warping, Scanning, Illumination, and Screen Photography.
• Capability Expansion: Now supports Seal Recognition, Text Spotting, and expands to 111 languages (including China’s Tibetan script and Bengali).
• Long Document Mastery: Supports automatic cross-page table merging and hierarchical heading identification.
• Try it now: Available on HuggingFace or our Official Website.

Released PaddleOCR-VL:

• Model Introduction:

  • PaddleOCR-VL is a SOTA and resource-efficient model tailored for document parsing. Its core component is PaddleOCR-VL-0.9B, a compact yet powerful vision-language model (VLM) that integrates a NaViT-style dynamic resolution visual encoder with the ERNIE-4.5-0.3B language model to enable accurate element recognition. This innovative model efficiently supports 109 languages and excels in recognizing complex elements (e.g., text, tables, formulas, and charts), while maintaining minimal resource consumption. Through comprehensive evaluations on widely used public benchmarks and in-house benchmarks, PaddleOCR-VL achieves SOTA performance in both page-level document parsing and element-level recognition. It significantly outperforms existing solutions, exhibits strong competitiveness against top-tier VLMs, and delivers fast inference speeds. These strengths make it highly suitable for practical deployment in real-world scenarios. The model has been released on HuggingFace. Everyone is welcome to download and use it! More introduction information can be found in PaddleOCR-VL.

• Core Features:

  • Compact yet Powerful VLM Architecture: We present a novel vision-language model that is specifically designed for resource-efficient inference, achieving outstanding performance in element recognition. By integrating a NaViT-style dynamic high-resolution visual encoder with the lightweight ERNIE-4.5-0.3B language model, we significantly enhance the model’s recognition capabilities and decoding efficiency. This integration maintains high accuracy while reducing computational demands, making it well-suited for efficient and practical document processing applications.
  • SOTA Performance on Document Parsing: PaddleOCR-VL achieves state-of-the-art performance in both page-level document parsing and element-level recognition. It significantly outperforms existing pipeline-based solutions and exhibiting strong competitiveness against leading vision-language models (VLMs) in document parsing. Moreover, it excels in recognizing complex document elements, such as text, tables, formulas, and charts, making it suitable for a wide range of challenging content types, including handwritten text and historical documents. This makes it highly versatile and suitable for a wide range of document types and scenarios.
  • Multilingual Support: PaddleOCR-VL Supports 109 languages, covering major global languages, including but not limited to Chinese, English, Japanese, Latin, and Korean, as well as languages with different scripts and structures, such as Russian (Cyrillic script), Arabic, Hindi (Devanagari script), and Thai. This broad language coverage substantially enhances the applicability of our system to multilingual and globalized document processing scenarios.

Released PP-OCRv5 Multilingual Recognition Model:

• Improved the accuracy and coverage of Latin script recognition; added support for Cyrillic, Arabic, Devanagari, Telugu, Tamil, and other language systems, covering recognition of 109 languages. The model has only 2M parameters, and the accuracy of some models has increased by over 40% compared to the previous generation.

Significant Model Additions:

• Introduced training, inference, and deployment for PP-OCRv5 recognition models in English, Thai, and Greek. The PP-OCRv5 English model delivers an 11% improvement in English scenarios compared to the main PP-OCRv5 model, with the Thai and Greek recognition models achieving accuracies of 82.68% and 89.28%, respectively.

Deployment Capability Upgrades:

• Full support for PaddlePaddle framework versions 3.1.0 and 3.1.1.
• Comprehensive upgrade of the PP-OCRv5 C++ local deployment solution, now supporting both Linux and Windows, with feature parity and identical accuracy to the Python implementation.
• High-performance inference now supports CUDA 12, and inference can be performed using either the Paddle Inference or ONNX Runtime backends.
• The high-stability service-oriented deployment solution is now fully open-sourced, allowing users to customize Docker images and SDKs as required.
• The high-stability service-oriented deployment solution also supports invocation via manually constructed HTTP requests, enabling client-side code development in any programming language.

Benchmark Support:

• All production lines now support fine-grained benchmarking, enabling measurement of end-to-end inference time as well as per-layer and per-module latency data to assist with performance analysis. Here's how to set up and use the benchmark feature.
• Documentation has been updated to include key metrics for commonly used configurations on mainstream hardware, such as inference latency and memory usage, providing deployment references for users.

Bug Fixes:

• Resolved the issue of failed log saving during model training.
• Upgraded the data augmentation component for formula models for compatibility with newer versions of the albumentations dependency, and fixed deadlock warnings when using the tokenizers package in multi-process scenarios.
• Fixed inconsistencies in switch behaviors (e.g., use_chart_parsing) in the PP-StructureV3 configuration files compared to other pipelines.

Other Enhancements:

• Separated core and optional dependencies. Only minimal core dependencies are required for basic text recognition; additional dependencies for document parsing and information extraction can be installed as needed.
• Enabled support for NVIDIA RTX 50 series graphics cards on Windows; users can refer to the installation guide for the corresponding PaddlePaddle framework versions.
• PP-OCR series models now support returning single-character coordinates.
• Added AIStudio, ModelScope, and other model download sources, allowing users to specify the source for model downloads.
• Added support for chart-to-table conversion via the PP-Chart2Table module.
• Optimized documentation descriptions to improve usability.