OCRRouter

SmolDocling vs Donut

Pick SmolDocling if you want to turn English pages into Markdown, HTML or JSON with bounding boxes and are only experimenting, since it is a preview release. Pick Donut if your documents are narrow and fixed-layout, such as receipts or forms, and you can fine-tune it to extract fields as JSON. The main split: SmolDocling converts whole pages, while Donut is a base model that does nothing useful until you train it.

SmolDoclinghuggingface/smoldoclingDonutnaver/donut
TypeOpen sourceOpen source
Price1 credit / page1 credit / page
Per 1,000 pages$1.00$1.00
Typical speed1.0 s / page1.0 s / page
Best forVery cheap page conversionReceipt and form fields after fine-tuning
InputsPDF, imagesimages
LanguagesMainly EnglishEnglish, Chinese, Japanese, Korean
Max pages per request500100
Tables✓ YesNo
HandwritingNoNo
Math✓ YesNo
Bounding boxes✓ YesNo
ConfidenceNoNo
FormsNo✓ Yes
MultilingualNoNo
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 SmolDocling if

  • It converts full pages and outputs DocTags, Markdown, HTML, JSON, LaTeX and bounding boxes.
  • It runs at about 0.35 seconds per page on an A100 with vLLM, and its 256M parameters fit modest hardware.
  • One model handles text, tables, code, formulas and charts with location tags, with no training step.

Choose Donut if

  • It is MIT licensed, and the Apache-2.0 vs MIT question does not come up with SmolDocling's mixed licence signals (Apache-2.0 in the summary, CDLA-Permissive-2.0 in the metadata tag).
  • After fine-tuning it extracts receipt and form fields straight to JSON, with no external OCR step.
  • A Hacker News practitioner reports that about 200 documents of fine-tuning can reach human-level accuracy on narrow data.

Key differences

SmolDoclingDonut
LicenceApache-2.0 per model summary; metadata tag says CDLA-Permissive-2.0MIT
Ready to useYes, but a preview release; authors point users to Granite-DoclingNo. The base checkpoint is pretrain-only and must be fine-tuned first
Output formatsDocTags, Markdown, HTML, JSON, LaTeX, bounding boxesJSON, plain text
Model size256M parametersNot published
BenchmarksTEDS 0.82 on FinTabNet tables, full-page OCR F1 0.80, OCRBench 338 (author-reported)Not published
LanguagesEnglish onlyNot published
Known failureEndless token repetition and missing closing tags; Docling issue #1515 reports output loops that make generation very slowMisses tiny text in large images because of input resolution limits; one HN user found results on personal documents did not match the promised scores
Setup caveatNot publishedTransformers v5 dropped image-to-text pipeline support, so load the model directly or use v4

Sources

What each one is

SmolDocling

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

SmolDocling details

Donut

OCR-free document understanding model. Fine-tuned for receipts, forms and document Q&A.

Donut 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": "huggingface/smoldocling","fallbacks": ["naver/donut"],"file_url": "https://example.com/document.pdf"}

Questions

More comparisons

SmolDocling alternativesDonut alternativesBest OCR models in 2026