PaddleOCR: An Industrial-Grade Engine for Turning Documents into Structured Data

Forum topic · 小凯 · 2026-07-03

Summary

PaddleOCR is an Apache 2.0 open-source OCR toolkit from PaddlePaddle with 84.6K GitHub stars, widely integrated into Dify, RAGFlow, Cherry Studio, and OmniParser as the default bridge from PDFs, images, and scans into LLM-ready JSON or Markdown. The 3.7.0 release (June 2026) ships three product lines: PP-OCRv6, a 50-language scene-text recognition model in tiny, small, and medium tiers (1.5M to 34.5M parameters) with 5.2x OpenVINO CPU acceleration and accuracy gains of 4.6% detection and 5.1% recognition over v5, reportedly beating Qwen3-VL-235B and GPT-5.5 on text tasks; PaddleOCR-VL-1.6, a 0.9B-parameter vision-language model that reaches 96.3% on OmniDocBench v1.6, combining a NaViT dynamic-resolution vision encoder with ERNIE-4.5-0.3B; and PP-StructureV3 for fine-grained layout, table-cell coordinates, and Office-to-Markdown conversion. The project also offers MCP, LangChain, browser, mobile, and ONNX/TensorRT/OpenVINO deployment paths, used by 6,500+ downstream projects for RAG ingestion, visual agents, and on-device document AI.

Key points

Project positioning. PaddleOCR (PaddlePaddle/PaddleOCR) is framed as document-AI infrastructure rather than a simple OCR tool, turning PDFs, images, and scans into LLM-ready structured data (JSON or Markdown).
Adoption. Integrated by Dify, RAGFlow, Cherry Studio, Pathway, OmniParser, and an official MCP server; referenced by 6,500+ downstream projects. Multi-language SDKs cover Go, TypeScript, Java, C++, C#, plus Android/iOS and a browser-side PaddleOCR.js.
Three product lines (release 3.7.0, 2026-06-11).
*PP-OCRv6* (scene text): 50-language unified model (Chinese, English, Japanese, 46 Latin-script languages) with no model switching. Tiers are tiny (1.5M params), small (7.7M), medium (34.5M). Reported CPU speedups: 5.2x with OpenVINO, 6.1x on Apple M4, ~0.13s on A100. Accuracy gains over v5 are +4.6% detection and +5.1% recognition; the post claims it surpasses Qwen3-VL-235B and GPT-5.5 on text-recognition benchmarks.
*PaddleOCR-VL-1.6* (document parsing VLM): 0.9B parameters built on a NaViT dynamic-resolution vision encoder plus ERNIE-4.5-0.3B language model. Reported SOTA on OmniDocBench v1.6 at 96.3%. Outputs Markdown or JSON, handles tables, formulas, seals, charts, ancient texts, and rare characters, with automatic cross-page table merging and hierarchical heading detection.
*PP-StructureV3* (structured extraction): fine-grained coordinates for table cells, text boxes, and regions; Word/Excel/PowerPoint-to-Markdown conversion; complements PaddleOCR-VL by focusing on precise layout rather than paragraph-level understanding.
Why it matters now. Three converging trends drive adoption: enterprise RAG pipelines need OCR for ~80% of inputs that are PDFs/scans, visual agents need structured document outputs, and privacy-sensitive edge deployment demands small models that run on CPU and mobile.
Architecture note. The VLM line and the traditional pipeline (PP-OCRv6 + PP-StructureV3) run in parallel rather than as replacements: VLMs optimize for semantic understanding, pipelines optimize for exact coordinates, fonts, and seal positions. The team's stated approach is full-stack coverage rather than a single bet on one route.

References

GitHub: https://github.com/PaddlePaddle/PaddleOCR
Official site: https://www.paddleocr.com
PaddleOCR-VL-1.6 paper: arXiv:2606.03264
PaddleOCR 3.0 technical report: arXiv:2507.05595

PaddleOCR: An Industrial-Grade Engine for Turning Documents into Structured Data

Summary

Key points

References

Tags