Source-linked AI summary
PaddleOCR 3.0 Technical Report
Cheng Cui, Ting Sun, Manhui Lin, Tingquan Gao, Yubo Zhang, Jiaxuan Liu, Xueqing Wang, Zelun Zhang, Changda Zhou, Hongen Liu, Yue Zhang, Wenyu Lv, Kui Huang, Yichao Zhang, Jing Zhang, Jun Zhang, Yi Liu, Dianhai Yu, Yanjun Ma
TL;DR
Modern LLM and RAG systems need OCR that can support accurate text processing, document understanding, and knowledge extraction across complex real-world materials. PaddleOCR 3.0 addresses this need with multilingual recognition, hierarchical parsing, semantic information extraction, and an open deployment-oriented toolkit. Its solutions are reported to provide broad, competitive capabilities across recognition and document parsing tasks while supporting research and industrial use.
Problem
Modern LLM and RAG systems require OCR that can process complex documents, diverse languages, and multimodal content while supporting dataset construction and knowledge extraction.
Method
PaddleOCR 3.0 combines PP-OCRv5 for multilingual recognition, PP-StructureV3 for structured document parsing, PP-ChatOCRv4 for key information extraction, and tools for training, inference, and deployment.
Results
PaddleOCR 3.0 provides accurate and efficient recognition and parsing solutions, with PP-StructureV3 reported as current SOTA on Chinese and English document parsing and competitive with expert and general VLMs.
Takeaways & Limitations
PaddleOCR 3.0 offers an open, extensible infrastructure for document AI spanning multilingual OCR, document parsing, semantic extraction, and deployment.
Abstract
from arXiv · showhide
This technical report introduces PaddleOCR 3.0, an Apache-licensed open-source toolkit for OCR and document parsing. To address the growing demand for document understanding in the era of large language models, PaddleOCR 3.0 presents three major solutions: (1) PP-OCRv5 for multilingual text recognition, (2) PP-StructureV3 for hierarchical document parsing, and (3) PP-ChatOCRv4 for key information extraction. Compared to mainstream vision-language models (VLMs), these models with fewer than 100 million parameters achieve competitive accuracy and efficiency, rivaling billion-parameter VLMs. In addition to offering a high-quality OCR model library, PaddleOCR 3.0 provides efficient tools for training, inference, and deployment, supports heterogeneous hardware acceleration, and enables developers to easily build intelligent document applications.
1. Introduction
PaddleOCR 3.0 responds to increasingly demanding OCR and document-understanding needs in modern AI by combining enhanced recognition, document parsing, and semantic extraction capabilities with an open, extensible toolkit.
- Motivation: Modern LLM and RAG systems require OCR that supports dataset construction, knowledge extraction, and links between visual and semantic information.OCR increasingly serves as infrastructure for digitizing scientific literature and enabling question answering over heterogeneous documents.
- Motivation: Existing OCR systems struggle with low-quality scans, complex layouts, diverse scripts, and the gap between research prototypes and production-ready document intelligence.Practical users also seek lightweight solutions that adapt to hardware constraints and integrate with large-scale AI workflows.
- Project evolution: PaddleOCR evolved from earlier open-source OCR iterations that targeted comprehensive coverage, end-to-end workflows, and lightweight efficiency.The report presents version 3.0 as a milestone upgrade building on the project’s prior development.
- Motivation: Community requirements increasingly emphasize handwriting recognition, multilingual and rare-script support, complex document parsing, and key information extraction.These needs are tied to extracting, structuring, and semantically interpreting diverse documents for LLM and RAG applications.
- PaddleOCR 3.0: PaddleOCR 3.0 introduces PP-OCRv5 for accurate multilingual recognition, PP-StructureV3 for integrated document parsing, and PP-ChatOCRv4 for semantic key information extraction.PP-OCRv5 covers printed, handwritten, and multilingual documents; PP-StructureV3 combines layout, table, and structure analysis; PP-ChatOCRv4 supports context-aware extraction and question answering.
- PaddleOCR 3.0: PaddleOCR 3.0 prioritizes openness, usability, extensibility, deployment integration, and backward compatibility while expanding specialized document-processing capabilities.The release includes a robust API ecosystem, model zoo, community support, and solutions for seal text, formulas, and charts.
2. Core Capabilities
PaddleOCR 3.0 provides three core capabilities: multilingual OCR, structured document parsing, and key information extraction. Its pipelines combine specialized modules and report strong performance across complex recognition and document-understanding scenarios.
- PP-OCRv5: PP-OCRv5 is a lightweight, high-precision OCR system supporting Simplified Chinese, Traditional Chinese, Chinese Pinyin, English, and Japanese in one model.Its unified multilingual architecture maintains a compact size under 100 MB.
- PP-OCRv5: 26% lower recognition error rates were achieved on non-standard Chinese and English handwriting tasks compared with previous models.The evaluation targets complex handwriting relevant to examination grading, bill recognition, and contract entry.
- PP-StructureV3: PP-StructureV3 converts document images or PDF files into structured JSON and Markdown through preprocessing, OCR, layout analysis, item recognition, and postprocessing.Its layout and recognition modules address complex documents, tables, formulas, charts, and diverse layouts.
- PP-StructureV3: PP-StructureV3 achieved state-of-the-art performance on Chinese and English document parsing in the OmniDocBench evaluation.It outperformed pipeline-based tools and remained competitive with expert and general VLMs.
- PP-ChatOCRv4: PP-ChatOCRv4 combines LLMs, VLMs, and OCR for key information extraction from challenging documents, including complex layouts, multi-page PDFs, tables, rare characters, and seals.PP-DocBee2 additionally supports prompt-based extraction of text answers directly from document images.
3. Codebase Architecture Design
PaddleOCR 3.0 reorganizes its inference library into a layered architecture built on PaddleX 3.0, while preserving backward compatibility and separating training from inference responsibilities.
- The codebase separates a model training toolkit from a lightweight, efficient inference library supporting official and user-trained models.
- PaddleOCR 3.0 upgrades its inference library on PaddleX 3.0 and considers backward compatibility to minimize migration effort from PaddleOCR 2.x.
- The redesigned inference library uses three layers: interface, foundation, and an intermediate layer described in the architecture.
- Unified Python APIs, redesigned task-specific CLI subcommands, and configurable deployment foundations improve usability and functional separation.
- The layered architecture promotes loose coupling, modularity, and ease of maintenance because higher-level components depend only on lower-level abstractions.
4. Deployment
PaddleOCR 3.0 provides flexible deployment across optimized inference, scalable serving, mobile devices, and MCP-compatible hosts, addressing latency, throughput, and hardware constraints.
- High-Performance Inference: 73.1% lower PP-OCRv5_mobile_rec latency and 40.4% lower PP-OCRv5_mobile_det latency are achieved on NVIDIA Tesla T4 devices with high-performance inference.
- High-Performance Inference: High-performance inference automatically selects suitable backends and applies optimizations such as multi-threading, FP16 inference, and model conversion.
- Serving: Pipeline serving offers Basic Serving for lightweight, low-concurrency validation and High-Stability Serving through Triton for demanding stability and performance requirements.
- Serving: Both serving solutions share similar interfaces, allowing users to begin with basic serving and migrate to high-stability serving without significant migration costs.
- On-Device Deployment: PaddleOCR 3.0 supports PP-OCR deployment on resource-constrained mobile platforms through model optimization and Paddle-Lite5 integration.
- MCP Server: Its lightweight MCP server exposes OCR and PP-StructureV3 pipelines to MCP-compatible hosts and supports multiple inference and deployment modes.
5. Conclusion
PaddleOCR 3.0 is presented as a milestone upgrade centered on PP-OCRv5, PP-StructureV3, and PP-ChatOCRv4, with further model expansion planned.
- PaddleOCR 3.0 is described as a milestone upgrade featuring PP-OCRv5, PP-StructureV3, and PP-ChatOCRv4.
- The authors state that future releases will expand multilingual text recognition, multimodal OCR, and document parsing models.
B. Usage of command and API details
PaddleOCR 3.0 can be installed from PyPI and used through either a command-line interface or a Python API.
- PaddleOCR 3.0 provides PyPI installation together with command-line and Python API access.
B.1. Run inference by CLI
The CLI examples show how to invoke PP-OCRv5, PP-StructureV3, and PP-ChatOCRv4 on an input image with pipeline-specific commands and options.
- B.1. Run inference by CLI: The CLI methods are provided to let users quickly experience PP-OCRv5, PP-StructureV3, and PP-ChatOCRv4.
- B.1. Run inference by CLI: PP-OCRv5 inference uses the `paddleocr ocr` command with an input image and document-processing options.The example disables document orientation classification, document unwarping, and textline orientation.
- B.1. Run inference by CLI: PP-StructureV3 inference uses the `paddleocr pp_structurev3` command with an input image.
- B.1. Run inference by CLI: PP-ChatOCRv4 inference uses the `paddleocr pp_chatocrv4_doc` command and requires a Qianfan API key.The example also disables document orientation classification and document unwarping.
B.2. Run inference by Python API
The Python API provides a clean interface for using and integrating PaddleOCR into users’ own projects.
- B.2. Run inference by Python API: PaddleOCR provides a clean Python interface for integrating its capabilities into users’ own projects.
1. PP-OCRv5 Example
The PP-OCRv5 Python example initializes the OCR class, runs inference on an image, and saves visual and JSON outputs.
- 1. PP-OCRv5 Example: The example initializes a `PaddleOCR` instance before running OCR inference.
- 1. PP-OCRv5 Example: OCR inference is run on a sample image after the `PaddleOCR` instance is initialized.
- 1. PP-OCRv5 Example: The example visualizes inference results and saves the JSON results.
- 1. PP-OCRv5 Example: Each result is saved to an output directory as an image and as JSON.
2. PP-StructureV3 Example
The examples initialize PP-StructureV3 and PP-ChatOCRv4 components, run document inference, collect visual information, and configure local or service-based MCP connections.
- 2. PP-StructureV3 Example: The example initializes a `PPStructureV3` instance and runs PP-StructureV3 inference.
- 2. PP-StructureV3 Example: PP-StructureV3 outputs are visualized and saved as JSON results.
- 3. PP-ChatOCRv4 Example: The PP-ChatOCRv4 example initializes `PPChatOCRv4Doc` and collects visual information from prediction results.
- 3. PP-ChatOCRv4 Example: The ChatOCR configuration specifies a chat-bot module with a Qianfan endpoint and API key, alongside a retriever using the `embedding-v1` model.
- MCP host configurations: The MCP host examples support local, AI Studio, and self-hosted PaddleOCR sources through the `paddleocr_mcp` command.
- MCP host configurations: MCP configurations select the OCR pipeline and provide source-specific settings such as access tokens or server URLs.