OCRRouter

PaddleOCR (PP-OCR) online

Baidu's classic open-source text detection and recognition pipeline. Light enough to run on CPU.

PaddleOCR (PP-OCR) versions: PP-OCRv5 (May 20, 2025, with PaddleOCR 3.0), PP-OCRv4 (September 2023, with PaddleOCR 2.7)

PDF, PNG, JPG, WebP, HEIC. Up to 20 MB each, 10 files.No file?
Model ID
paddle/paddleocr
Type
Open source
Price
1 credit / page ($1.00 per 1,000 pages)
Typical speed
0.7 s / page
Inputs
PDF, images
Languages
80+ languages
Max pages per request
500
Output
Markdown, JSON with positions
License
Apache-2.0
Supports
Bounding boxesConfidenceMultilingual
curl https://api.ocrrouter.ai/v1/ocr \
  -H "Authorization: Bearer $OCRROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "paddle/paddleocr",
    "file_url": "https://example.com/invoice.pdf",
    "format": "markdown"
  }'

PaddleOCR (PP-OCR) speed and reliability

Loading…

Measured on our own traffic over the last 7 days.

PaddleOCR (PP-OCR) pricing

PaddleOCR (PP-OCR) costs 1 credits per page. One dollar buys 1,000 credits, so 1,000 pages cost $1.00. You only pay for pages that return a result.

New accounts get 50 free credits, enough for about 50 pages with this model. See pricing for credit packs.

Similar models

All PaddleOCR (PP-OCR) alternativesBest OCR models by use case

Compare PaddleOCR (PP-OCR)

PaddleOCR (PP-OCR) versions

What is PaddleOCR?

PaddleOCR is Baidu's open-source OCR toolkit, built on the PaddlePaddle deep-learning framework, with over 70,000 GitHub stars. The name covers several product lines. PP-OCR (this page) is the classic two-stage pipeline: a text detection model finds text lines, and a recognition model reads them. Optional modules fix document orientation and straighten text lines. PP-StructureV3 adds layout, table and formula parsing to Markdown or JSON. PaddleOCR-VL is a separate 0.9B vision-language model for full document parsing. PP-OCR models are small (PP-OCRv6 tiers run from 1.5M to 34.5M parameters), so they run on CPU, mobile and edge devices, and return text lines with box coordinates and confidence scores.

Install and run PaddleOCR in Python

PaddleOCR 3.x supports Python 3.8 to 3.12 on Linux, Windows and macOS. Install the PaddlePaddle framework first (CPU or GPU build, see the official install guide), then the package:

python -m pip install paddlepaddle   # CPU build
python -m pip install paddleocr
paddleocr ocr -i scan.png              # CLI
from paddleocr import PaddleOCR

ocr = PaddleOCR(use_doc_orientation_classify=False,
                use_doc_unwarping=False,
                use_textline_orientation=False)
result = ocr.predict('scan.png')
for res in result:
    res.print()               # text, boxes, scores
    res.save_to_json('output')

Models download on first run. Since 3.5, many models can also run on a Transformers backend, and ONNX export with OpenVINO, TensorRT or ONNX Runtime is documented.

PP-OCR versions and languages

PP-OCRv5 (May 2025, PaddleOCR 3.0) reads Simplified and Traditional Chinese, Pinyin, English and Japanese in one model. Its multilingual recognition models later grew to 109 languages, including Cyrillic, Arabic, Devanagari, Telugu and Tamil. PP-OCRv6 (PaddleOCR 3.7, June 2026) covers 50 languages with one unified model (Chinese, English, Japanese and 46 Latin-script languages). It comes in tiny, small and medium tiers. PaddleOCR reports +4.6% detection and +5.1% recognition accuracy over PP-OCRv5 server, and a 5.2x CPU speedup with OpenVINO. Across all its models the project says it supports more than 100 languages. Each language family may need its own recognition model, set with the lang argument.

PaddleOCR license (Apache 2.0)

PaddleOCR's code and released models are under the Apache License 2.0, as stated in the LICENSE file and README badge. You can use, modify and ship it in commercial products. You must keep the license and copyright notices and mark files you changed, and you get an explicit patent grant from contributors. There is no copyleft requirement, so your own code can stay closed. Dependencies, including PaddlePaddle (also Apache-2.0), have their own licenses, and any third-party datasets you use for fine-tuning keep their own terms.

PaddleOCR vs EasyOCR vs Tesseract

All three are free and open source. Tesseract (Apache-2.0) is a C++ engine with LSTM line recognition. It installs easily, runs fast on CPU and handles clean printed scans well, but it has no learned text detector, so photos, rotated text and complex layouts are its weak spots. EasyOCR (Apache-2.0) is a PyTorch library (CRAFT detector plus CRNN recognizer) with a very simple API and about 80 languages. PaddleOCR has a learned detector and recognizer, is strongest on Chinese and mixed Chinese-English text, handles scene text and skewed lines well, and adds layout, table and document parsing through PP-StructureV3 and PaddleOCR-VL. It needs the PaddlePaddle framework or an ONNX or Transformers backend. In practice: Tesseract for clean scans, PaddleOCR for accuracy and CJK, EasyOCR for quick Python prototypes.

Why PaddleOCR can feel slower than EasyOCR

Speed reports differ because the two libraries do different amounts of work by default. A PaddleOCR 3.x pipeline can run document orientation classification, document unwarping and text-line orientation before detection and recognition. Turning these off with use_doc_orientation_classify=False, use_doc_unwarping=False and use_textline_orientation=False removes whole model passes. The server-tier models are larger than the mobile or tiny tiers, so pick a smaller model for CPU. The first call also downloads and initializes models, and a CPU-only PaddlePaddle build without MKL-DNN or OpenVINO runs much slower than an accelerated one. PaddleOCR's high-performance inference docs cover ONNX Runtime, OpenVINO and TensorRT backends.

PaddleOCR on OCRRouter: status and live alternatives

PP-OCR is not live on OCRRouter yet. The API page and the price shown are placeholders until the adapter ships, and calls to paddle/paddleocr will not return results today. If you need it now, self-host it with the Python package or Baidu's official serving deployment. Live options on OCRRouter through the same endpoint: paddle/paddleocr-vl (PaddleOCR-VL from the same team, with layout, tables and formulas in 109 languages), deepseek/deepseek-ocr (Markdown output from PDFs and images) and tesseract/tesseract (free, for clean printed text).

Questions