OCRRouter

RapidOCR vs Surya

Pick RapidOCR for offline CPU OCR of Chinese and English screenshots and receipts, especially inside Java or C# apps. Pick Surya when you need layout, reading order, tables and bounding boxes across 90+ languages, and you can run an inference backend. The main split is plain text lines versus structured document parsing.

RapidOCRrapidai/rapidocrSuryadatalab/surya
TypeOpen sourceOpen source
Price1 credit / page2 credits / page
Per 1,000 pages$1.00$2.00
Typical speed0.6 s / page1.5 s / page
Best forLightweight OCR on CPU or edgeLayout and line-level OCR
InputsimagesPDF, images
LanguagesChinese, English and more90+ languages
Max pages per request100500
TablesNo✓ Yes
HandwritingNoNo
MathNoNo
Bounding boxes✓ Yes✓ Yes
Confidence✓ YesNo
FormsNoNo
Multilingual✓ Yes✓ Yes
On OCRRouterAvailableAvailable

Prices are what you pay on OCRRouter. Speed is the provider's typical figure, not a benchmark we ran.

Which to choose

Choose RapidOCR if

  • It installs with pip install rapidocr onnxruntime and runs PaddleOCR models on CPU with no Paddle framework or separate server.
  • It has bindings for Python, C++, Java and C#, so it embeds in desktop and server apps in many stacks.
  • Its licence is Apache-2.0 with no revenue cap, though converted models keep their upstream licences.

Choose Surya if

  • One model returns layout labels, reading order, OCR, HTML tables and KaTeX math with per-block bounding boxes.
  • It covers 90+ languages, while RapidOCR is Chinese and English only by default.
  • It scores 83.7% on olmOCR-bench (reported by the authors) with 689M parameters, and runs on a laptop CPU or Apple Silicon through llama.cpp.

Key differences

RapidOCRSurya
How you run itLibrary. pip install rapidocr onnxruntime, runs in-process.Lightweight client that needs a separate vLLM or llama.cpp inference server.
LicenceApache-2.0. Converted models keep their upstream licences.Code Apache-2.0. Weights under a modified AI Pubs Open Rail-M: free for research, personal use and companies under $5M funding or revenue.
HardwareCPU via ONNX Runtime by default. Optional GPU, TensorRT, OpenVINO, MNN, Paddle or PyTorch engines.NVIDIA GPU via vLLM, or CPU and Apple Silicon via llama.cpp. Peak memory about 4.3-5.9 GB on an M4 Max.
Output formatsPlain text and bounding boxes.JSON, HTML, LaTeX, bounding boxes and plain text.
LanguagesChinese and English by default. Other languages need extra models.90+ languages. Only 21 of 91 reach a 90% pass rate, and multilingual tables score 58.1.
Tables and layoutNot supported. Not meant for layout, tables or Markdown output.Layout detection, reading order and HTML tables. Surya 2.5 table output lacks row and column geometry.
Known failure16-bit grayscale PNG or TIFF give no text or wrong text. Arabic crashes because python-bidi is undeclared. Sliced inference crashes on large pages.Handwriting, complex tables and math are better sent to the hosted Datalab API. Not for photos or natural scenes. Apple M4 Max throughput is 0.198 pages/s at 96 DPI with llama.cpp Q8_0.

Sources

What each one is

RapidOCR

PaddleOCR models packaged for ONNX Runtime and other engines. Easy to deploy without Paddle.

RapidOCR details

Surya

Open-source OCR toolkit with line detection, layout analysis, reading order and tables.

Surya details

Try both on your own file

Run the file once with each model and compare the output.

PDF, PNG, JPG, WebP, HEIC. Up to 20 MB each, 10 files.No file?

Use one, fall back to the other

{"model": "rapidai/rapidocr","fallbacks": ["datalab/surya"],"file_url": "https://example.com/document.pdf"}

Questions

More comparisons

RapidOCR alternativesSurya alternativesBest OCR models in 2026