PaddleOCR: Turn Images and PDFs Into Structured Text (GitHub, Scanned)

PaddleOCR reads text out of images and PDFs. At its simplest, one command finds every line of text in a photo or scan and returns the words with their positions; the PP-OCRv6 models handle 50 languages in a single model, and the toolkit covers more than 100 in all, along with street signs, ID cards, dot-matrix print and other hard cases. Beyond plain OCR, PP-StructureV3 and the PaddleOCR-VL vision-language models parse whole documents, including tables, formulas, charts and seals, into Markdown or JSON ready to feed to a language model.

It is on this list because it has become plumbing for AI apps: Dify, RAGFlow and Cherry Studio use it to get documents into retrieval systems. The models are small, from 1.5 million parameters for the tiny OCR model to 0.9 billion for PaddleOCR-VL, so they run on a CPU or edge device, and they can be exported to ONNX or run through OpenVINO, TensorRT or Hugging Face Transformers instead of the PaddlePaddle framework.

PaddleOCR is developed by the PaddlePaddle team at Baidu, has about 90,000 stars and is licensed Apache-2.0. Version 3.7.0, with PP-OCRv6, was released in June 2026.

• Repository: github.com/PaddlePaddle/PaddleOCR
• Licence: Apache-2.0 (Apache License 2.0)
• Language: Python. Stars: 90.5K. Forks: 11.4K. Last push: Sep 16, 2026.
• Scan: safe, Sep 16, 2026, commit dab3fe3

Who it is for

Developers building document pipelines, RAG systems or data-entry automation, and anyone who needs accurate OCR for languages beyond English, especially Chinese and Japanese.

Getting started

1. Install the PaddlePaddle framework, CPU build shown (the docs list CUDA builds)

python -m pip install paddlepaddle==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cpu/

2. Install PaddleOCR (use "paddleocr[all]" for document parsing and translation too)

python -m pip install paddleocr

3. Read the text in a sample image and save the results

paddleocr ocr -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_002.png --save_path ./output

4. Parse a whole document page into Markdown and JSON with PP-StructureV3

paddleocr pp_structurev3 -i ./your_document.png

Models download on first use. To try it without installing anything, paddleocr.com has an online Experience Center and APIs.

Safety scan

We cloned PaddlePaddle/PaddleOCR at commit dab3fe3 on Sep 16, 2026 and ran the checks described on the GitHub Tools page: credential patterns, decode-and-execute code, install-time scripts, committed binaries, risky CI workflows, every host the code talks to, known vulnerabilities in pinned dependencies, and project hygiene. A person read every hit. This is what we found.

• No secrets, no pattern hits, no bare-IP URLs and no install hooks across 2,455 files and about 359,000 lines (121,000 Python, 150,000 Markdown). The two binaries are Gradle wrapper jars in the Android demo apps. The installer scripts are demos; the Arm Virtual Hardware one runs sudo pip install and a sudo setup script on that virtual board.
• Network hosts are Baidu's object storage (paddleocr.bj.bcebos.com and paddle-model-ecology.bj.bcebos.com) for model weights and demo images, GitHub, and Apache licence links in file headers; we found no analytics.
• The paddleocr package has no lockfile, so its own dependencies were not checked. The 131 advisories (3 critical, 60 high) are in langchain-paddleocr/uv.lock (73, including anyio, langchain-core, langsmith and pillow), paddleocr-js (49) and api_sdk/typescript (9, build tooling). In the browser SDK, protobufjs 7.5.4 (critical) is a runtime dependency through onnxruntime-web, which uses it to read model files; vitest's critical is test tooling.
• 9 workflows, none using pull_request_target; all 4 third-party actions are pinned to commits. Dependabot and licence present; no security policy, contributing guide or CodeQL. OpenSSF Scorecard 4.2.

What the scanner counted

The raw findings

Every hit the scanner wrote out, with a link to the exact line at the scanned commit. Secrets candidates are redacted.

By the numbers

Lines by language: Markdown 149.6K, Python 121K, YAML 28.5K, C++ 20.8K, TypeScript 11.9K, Shell 6,774.

Questions

Is PaddleOCR free?

Yes. PaddleOCR and its models are Apache-2.0, free for commercial use. Baidu also offers an online Experience Center and API at paddleocr.com for people who would rather not run it themselves.

Does PaddleOCR need a GPU?

No. The PP-OCR models are small enough for a CPU; the 3.7.0 release notes cite a 5.2 times CPU speedup with OpenVINO. A GPU helps for large batches and for the PaddleOCR-VL document model.

Can I use PaddleOCR without the PaddlePaddle framework?

Yes. Since version 3.5 you can choose the inference engine: PaddlePaddle, Hugging Face Transformers (for about 20 major models) or ONNX Runtime with exported models. The docs recommend installing only one engine per environment, since their dependencies can conflict.

This post is part of GitHub Tools, where every repository is cloned and scanned before it is written up. The scan is a snapshot of one commit on one day; the repository has moved on since, so check it before you install.