PaddleOCR-VL

What is PaddleOCR-VL?

PaddleOCR-VL is a lightweight multimodal released by Baidudocument resolutionmodel, designed for complex document structure parsing, the core parameters of only 0.9B, but with 92.6 points on top of the global authoritative document parsing evaluation list OmniBenchDoc V1.5, in the text, tables, formulas, reading order of the four core competencies of the four mainstream models comprehensively beyond the GPT-4o, Gemini-2.5 Pro, refreshing the performance of the global OCR VL models Ceiling. As a derivative model of Wenxin 4.5, it integrates NaViT Dynamic Resolution Visual Coder and ERNIE-4.5-0.3B language model, balancing accuracy and efficiency, and supports 109 languages, covering multilingual scenarios such as Chinese, English, French, and Arabic.

Key Features of PaddleOCR-VL

1. Multilingual Text Recognition
   • be in favor of 109 languagesThe OCR system can recognize handwriting, vertical text, art fonts, and other complex forms, breaking the limitations of traditional OCR that only recognizes print.
   • Examples: double-column typesetting in academic papers, mixed multilingual texts, and handwritten manuscripts from historical archives can all be accurately recognized.
2. Complex Element Analysis
   • form recognition: Accurately analyze nested tables and merged cells in financial and statistical reports, support for OTSL format Output, Structured Efficiency Improvement 50%.
   • formula recognition: CDM scores up to 91.43It supports the generation of LaTeX format to restore complex mathematical formulas in papers and textbooks.
   • Graphical understanding: Convert visual data such as bar charts, line graphs, pie charts, etc. into structured tables to support automated analysis.
3. Layout analysis and reading order prediction
   • pass (a bill or inspection etc) PP-DocLayoutV2 The model localizes semantic regions (e.g., headlines, body text, pictures, figure notes) and predicts reading order with an error value of only 0.043, automatically restores human reading habits.
   • Examples: layout of a two-column academic paper, logical ordering of contract terms and conditions.
4. Structured Output
   • be in favor of Markdown cap (a poem) JSON Output in a format that preserves the document hierarchy (e.g., headings, lists, code blocks) for database storage, API return, or knowledge base construction.

Scenarios for the use of PaddleOCR-VL

• Government and Enterprise Document Management
  Automate the digitization of contracts, statements, and files, extracting key terms, amounts, dates, and other information to reduce manual entry errors.
• Research Information Extraction
  Parsing experimental data, references, and graphical information in academic papers to support researchers to quickly locate the core content.
• Education Applications
  Homework correction, formula recognition, chart analysis, assisting teachers to efficiently deal with handwritten content in students' homework.
• Intelligent Knowledge Base Construction
  Scan copies,PDF Convert to structured data to provide high-quality knowledge inputs for RAG (Retrieval-Augmented Generation) systems, thereby enhancing the accuracy of large language model responses.
• Cross-Language Document Processing
  Supports automatic parsing of multilingual documents for the knowledge management needs of internationalized enterprises.

PaddleOCR-VL's project address

• Project website::https://ernie.baidu.com/blog/zh/posts/paddleocr-vl/
• HuggingFace Model Library::https://huggingface.co/PaddlePaddle/PaddleOCR-VL
• arXiv Technical Paper::https://arxiv.org/pdf/2510.14528
• Online Experience Demo::https://huggingface.co/spaces/PaddlePaddle/PaddleOCR-VL_Online_Demo
• Official Experience Address::https://aistudio.baidu.com/application/detail/98365

Recommended Reasons

• superior performance
  The world's best overall performance in OmniDocBench V1.5, with text editing distances of just 0.035TEDS Score 93.52, far superior to similar models.
• Lightweight and efficient
  Core parameters 0.9BThe speed of reasoning is up to 1881 token/s(single A100 GPU), up 14.2% from MinerU2.5, suitable for edge device deployment.
• Strong multimodal understanding
  Breaking through the limitations of traditional OCR, it realizes the ability to "read and understand documents", and supports complex layout analysis, handwriting recognition, and structured conversion of charts and diagrams.
• Open source and ecological compatibility
  Fully open source, support HuggingFace and GitHub platform, can be deeply integrated with the RAG system, and become the key infrastructure for AI knowledge processing.
• Wide range of scene coverage
  Applicable to government and enterprises, scientific research, education, knowledge management and other fields, to meet the globalization of document processing needs.