OCRRouter

PaddleOCR-VL vs SmolDocling

Pick PaddleOCR-VL for multilingual documents, formulas, tables and charts that you run on your own GPU. Pick SmolDocling only for English experiments on very small hardware, since its authors point users to Granite-Docling for updates and support. The main gap is scope: PaddleOCR-VL covers 109 languages and has a higher published benchmark score, while SmolDocling is English only and prone to output loops.

PaddleOCR-VLpaddle/paddleocr-vlSmolDoclinghuggingface/smoldocling
TypeOpen sourceOpen source
Price6 credits / page1 credit / page
Per 1,000 pages$6.00$1.00
Typical speed1.3 s / page1.0 s / page
Best forCJK documentsVery cheap page conversion
InputsPDF, imagesPDF, images
Languages109 languagesMainly English
Max pages per request500500
Tables✓ Yes✓ Yes
HandwritingNoNo
Math✓ Yes✓ Yes
Bounding boxes✓ Yes✓ Yes
Confidence✓ YesNo
FormsNoNo
Multilingual✓ YesNo
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

  • You have scanned documents in languages other than English, including Cyrillic, Arabic, Devanagari and Thai scripts, since it supports 109 languages in one model.
  • You parse academic papers with formulas or reports with tables and charts, and want Markdown or JSON for full pages.
  • You want an independently measured score: olmOCR-Bench lists it at 80.0 overall, and its authors report 96.33 on OmniDocBench v1.6 for version 1.6.

Choose SmolDocling if

  • You are experimenting with tiny document VLMs and want a 256M model, which is under a third of PaddleOCR-VL's 0.9B.
  • You convert English pages and want DocTags output for Docling, plus Markdown, HTML, JSON, LaTeX and bounding boxes.
  • You need speed on a data-centre GPU: about 0.35 seconds per page on an A100 with vLLM.

Key differences

PaddleOCR-VLSmolDocling
Parameters0.9B256M
LicenceApache-2.0Apache-2.0 per model summary (metadata tag says CDLA-Permissive-2.0)
Languages109, including Cyrillic, Arabic, Devanagari and ThaiEnglish only
Output formatsMarkdown, JSONDocTags, Markdown, HTML, JSON, LaTeX, bounding boxes
HardwareNVIDIA GPU recommended; CPU possible but slow; no minimum VRAM statedSmall enough for modest hardware; no minimum stated
BenchmarksOmniDocBench v1.6 96.33 (authors, version 1.6); olmOCR-Bench 80.0 (independent)Table TEDS 0.82 on FinTabNet 150dpi, full-page OCR F1 0.80, OCRBench 338 (authors)
Known failureVersion 1.6 regresses Russian and Kazakh OCR on born-digital PDFs, decodes about 11x slower with use_cache false, and hangs on blurry imagesEndless token repetition and missing closing tags; output loops on some images, making generation very slow

Sources

What each one is

PaddleOCR-VL

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

PaddleOCR-VL details

SmolDocling

Tiny 256M vision model from Hugging Face and IBM that outputs DocTags. Predecessor of Granite Docling.

SmolDocling 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": ["huggingface/smoldocling"],"file_url": "https://example.com/document.pdf"}

Questions

More comparisons

PaddleOCR-VL alternativesSmolDocling alternativesBest OCR models in 2026