OCRRouter

Qianfan OCR vs Donut

Pick Qianfan OCR for full-page parsing of multilingual documents with tables, formulas and handwriting, where you want Markdown, JSON or HTML out of the box. Pick Donut for narrow, fixed-layout jobs such as receipts and forms, where you can fine-tune on your own labelled samples. The main split is that Qianfan OCR works without training and Donut does not.

Qianfan OCRbaidu/qianfan-ocrDonutnaver/donut
TypeOCR APIOpen source
Price2 credits / page1 credit / page
Per 1,000 pages$2.00$1.00
Typical speed2.0 s / page1.0 s / page
Best forChinese documentsReceipt and form fields after fine-tuning
InputsPDF, imagesimages
LanguagesChinese, English and moreEnglish, Chinese, Japanese, Korean
Max pages per request200100
Tables✓ YesNo
Handwriting✓ YesNo
Math✓ YesNo
Bounding boxesNoNo
ConfidenceNoNo
FormsNo✓ Yes
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 Qianfan OCR if

  • You need full-page output in Markdown, JSON, HTML, LaTeX or bounding boxes, and Donut lists only JSON and plain text.
  • You process many languages, handwriting or scene text, since Qianfan OCR covers 192 languages.
  • You want a published benchmark: 93.12 on OmniDocBench v1.5, plus 79.8 on olmOCR-Bench and 880 on OCRBench.

Choose Donut if

  • You have a narrow, fixed-layout document type such as receipts or forms and can fine-tune a model for it.
  • You want a permissive MIT licence and loading through Hugging Face Transformers, with no external OCR step.
  • You want a base model to adapt for classification or VQA, not only parsing.

Key differences

Qianfan OCRDonut
Ready to useUsed directly for document parsing; also offered hosted on the Baidu Qianfan platformBase checkpoint is pretrain-only and must be fine-tuned before use
LicenceApache-2.0, allows commercial useMIT
Output formatsMarkdown, JSON, HTML, LaTeX, bounding boxesJSON, plain text
BenchmarksOmniDocBench v1.5 overall 93.12; olmOCR-Bench 79.8; OCRBench 880 (author-reported)Not published
Languages and content192 languages, handwriting and scene textNot published
Hardware and speed4B parameters; benchmarked on one A100 with vLLM at 1.024 pages per second with W8A8; not suited to CPU-only deploymentNot published
Known failureThinking mode can emit up to 16,384 tokens, adding latencyMisses tiny text in large images because of input resolution limits; Transformers v5 dropped image-to-text pipeline support, so use direct loading or v4

Sources

What each one is

Qianfan OCR

Baidu's OCR-tuned vision model. Strong on Chinese documents and tables.

Qianfan OCR 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": "baidu/qianfan-ocr","fallbacks": ["naver/donut"],"file_url": "https://example.com/document.pdf"}

Questions

More comparisons

Qianfan OCR alternativesDonut alternativesBest OCR models in 2026