About paddleocr.dev
An independent, hands-on guide to PaddleOCR. We aren't Baidu or the PaddleOCR team — we're developers who use OCR and wanted one place where every command works and every number can be traced to a source or a test.
Not the official site. paddleocr.dev has no affiliation with Baidu, PaddlePaddle or the PaddleOCR maintainers. The official project is github.com/PaddlePaddle/PaddleOCR, with documentation at paddleocr.ai and a hosted service at paddleocr.com.
Why this site exists#
PaddleOCR is one of the most capable open-source OCR toolkits, but getting it running is harder than it should be. The 3.x rewrite broke most tutorials on the web. The install depends on an engine you choose yourself, from an index that isn't PyPI if you want a GPU. The official docs are extensive, but they sometimes disagree with each other — and occasionally with the code.
We built this site to close that gap: practical guides, tested commands, honest comparisons and tools that do the fiddly parts for you.
How we research and verify#
Primary sources first#
Facts come from the PaddleOCR source code and official documentation (we read the repository itself, not summaries), package metadata on PyPI, npm and Hugging Face, and each competing project's own documentation. Every page lists its sources at the bottom.
We run it#
We install PaddleOCR on a clean machine and run what we write about. That's how we found that an unpinned
pip install paddlepaddlecrashes PaddleOCR 3.7.0 on CPU, that PP-StructureV3 still uses PP-OCRv5 text models by default, and that ONNX Runtime was the fastest engine on our hardware. The full benchmark, test images and scripts are published so anyone can reproduce them.We label who said what#
Numbers published by Baidu are labelled as vendor-reported. Numbers we measured say so, with the hardware. Third-party benchmarks name their publisher. When sources conflict, we show both instead of picking the convenient one.
We don't fill gaps with guesses#
If something can't be verified — a price that isn't published, a speed we couldn't measure without a GPU — we say so, or leave it out.
We date everything#
Each page shows when it was last checked and which PaddleOCR version it reflects (currently 3.7.0). When a new release lands, pages are re-checked before their dates change.
We use software tools, including AI assistants, to help research, draft and check pages. That doesn't replace verification: every published claim is checked against its primary source or against our own test runs.
What we've tested ourselves#
| Area | What we ran |
|---|---|
| Installation | PaddleOCR 3.7.0 with PaddlePaddle 3.3.1 (fails), 3.2.0 (works) and ONNX Runtime only (works), Python 3.13, Linux. |
| Text OCR | PP-OCRv6 medium/small/tiny and PP-OCRv5 server/mobile on five documents with known text, on two engines. |
| Document parsing | PP-StructureV3 on an invoice and a paper page, default vs PP-OCRv6 text models. |
| Browser OCR | PaddleOCR.js 0.4.2 in Chromium, model loading from Baidu's CDN and Hugging Face — the basis of our online tool. |
| Not tested | PaddleOCR-VL and GPU throughput (we don't have the hardware); Intel Macs; ARM boards. Pages that cover these rely on official documentation and say so. |
Corrections#
If something on this site is wrong, outdated or unclear, email hello@paddleocr.dev with the page URL. We fix confirmed errors promptly and update the page's date. Significant corrections are noted on the page.
Independence and funding#
- We aren't paid by Baidu, PaddlePaddle or any OCR vendor, and no vendor reviews our content.
- The site doesn't currently show ads. If that changes, ads will never influence which tool we recommend, and our privacy policy will be updated first.
- Links to other projects are there because they help readers; none are paid or affiliate links.
Contact#
Questions, corrections, or something you'd like us to test: hello@paddleocr.dev.
For bugs in PaddleOCR itself, the right place is the official GitHub issue tracker — we can't fix the library, but we're happy to point you in the right direction.
Reusing our content#
You're welcome to quote our pages and reuse our benchmark data and test images with a link back to the source page. For authoritative answers about PaddleOCR's behaviour, cite the official repository as well.