OCRRouter

PaddleOCR-VL vs Surya

Pick PaddleOCR-VL for multilingual scans, academic papers with formulas, and report tables and charts on a GPU you host. Pick Surya when you need layout labels, reading order and per-block bounding boxes, or want to run locally on a laptop or Apple Silicon. Check the licence first: Surya's weights need a commercial licence above $5M revenue or funding, while PaddleOCR-VL is Apache-2.0.

PaddleOCR-VLpaddle/paddleocr-vlSuryadatalab/surya
TypeOpen sourceOpen source
Price6 credits / page2 credits / page
Per 1,000 pages$6.00$2.00
Typical speed1.3 s / page1.5 s / page
Best forCJK documentsLayout and line-level OCR
InputsPDF, imagesPDF, images
Languages109 languages90+ languages
Max pages per request500500
Tables✓ Yes✓ Yes
HandwritingNoNo
Math✓ YesNo
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 PaddleOCR-VL if

  • It is Apache-2.0 with no revenue cap, so there is no commercial licence to buy at any company size.
  • It covers 109 languages in one model, including Cyrillic, Arabic, Devanagari and Thai scripts.
  • It reads text, tables, formulas and charts and returns Markdown and JSON for full pages, and version 1.6 scores 96.33 on OmniDocBench v1.6.

Choose Surya if

  • It returns layout labels, reading order, OCR, HTML tables and KaTeX math with per-block bounding boxes from one model.
  • It runs on CPU and Apple Silicon through llama.cpp, with peak memory of about 4.3-5.9 GB on an M4 Max.
  • It is fast on a server GPU, at 4.52 pages/s at 192 DPI on one H100 with vLLM, and scores 83.7% on olmOCR-bench.

Key differences

PaddleOCR-VLSurya
LicenceApache-2.0Code Apache-2.0; weights under a modified AI Pubs Open Rail-M. Free for research, personal use and companies under $5M funding or revenue.
How you run itSelf-host via PaddleOCR, vLLM or transformers. The transformers path only does element-level recognition, not full page parsing.Self-host as a lightweight client that needs a separately running vLLM or llama.cpp inference server.
Size and hardware0.9B parameters. NVIDIA GPU recommended; CPU is possible but slow. No minimum VRAM is stated.689M parameters. NVIDIA GPU via vLLM, or CPU and Apple Silicon via llama.cpp.
Output formatsMarkdown, JSONJSON, HTML, LaTeX, bounding boxes, plain text
BenchmarksolmOCR-Bench 80.0 (independent); OmniDocBench v1.6 96.33 for version 1.6 (authors)olmOCR-bench 83.7% and OmniParseBench 85.9 for Surya 2.5 (authors)
Languages109 languages in one model90+ languages; 74.9% pass rate across 91 languages, and only 21 of 91 reach 90%
TablesRecognizes tables and charts. Users report missing table content in layout analysis on version 1.6.HTML tables, but Surya 2.5 output lacks row and column geometry. Multilingual tables score 58.1.
Known failureVersion 1.6 regresses Russian and Kazakh Cyrillic on born-digital PDFs, hangs on blurry images, and decodes about 11x slower because use_cache ships as false.Not for handwriting, complex multilingual tables or photos and natural scenes. The README sends hard documents to the hosted Datalab API.

Sources

What each one is

PaddleOCR-VL

Open-source document model from Baidu. Excellent on Chinese, Japanese and Korean.

PaddleOCR-VL 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": "paddle/paddleocr-vl","fallbacks": ["datalab/surya"],"file_url": "https://example.com/document.pdf"}

Questions

More comparisons

PaddleOCR-VL alternativesSurya alternativesBest OCR models in 2026