OCRRouter

Qianfan OCR vs Kraken

Pick Qianfan OCR for parsing modern documents with tables, formulas and handwriting in 192 languages, on a GPU you control. Pick Kraken for historical manuscripts, early prints and right-to-left scripts, where you can train layout and recognition on your own material. The main split is the material: Qianfan OCR is a ready-made 4B parsing model, while Kraken is a trainable toolkit.

Qianfan OCRbaidu/qianfan-ocrKrakenkraken/kraken
TypeOCR APIOpen source
Price2 credits / page1 credit / page
Per 1,000 pages$2.00$1.00
Typical speed2.0 s / page1.5 s / page
Best forChinese documentsHistorical documents and manuscripts
InputsPDF, imagesimages
LanguagesChinese, English and moreMany scripts, incl. right-to-left
Max pages per request200500
Tables✓ YesNo
Handwriting✓ Yes✓ Yes
Math✓ YesNo
Bounding boxesNo✓ Yes
ConfidenceNoNo
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 Qianfan OCR if

  • You parse modern documents and want Markdown, JSON, HTML, LaTeX or bounding boxes in one model.
  • You need tables, formulas, key information extraction or handwriting across 192 languages without training anything.
  • You want a published benchmark: 93.12 on OmniDocBench v1.5, 79.8 on olmOCR-Bench and 880 on OCRBench.

Choose Kraken if

  • You transcribe historical manuscripts or early prints and want to train layout analysis, reading order and recognition on your own material.
  • You work with Arabic, Hebrew or other right-to-left, BiDi or top-to-bottom scripts.
  • You feed digital humanities tools that expect ALTO, PageXML, abbyyXML or hOCR, and you want community-trained models from the public Zenodo repository.

Key differences

Qianfan OCRKraken
Built forEnd-to-end document parsing, key information extraction, tables and formulasHistorical documents and non-Latin scripts, with trainable layout and recognition
LicenceApache-2.0, commercial use allowedApache-2.0
How you run itOpen weights on Hugging Face, or hosted on the Baidu Qianfan platform. Benchmarked on one NVIDIA A100 with vLLM.Python library installed with pip on Linux or macOS, x64 or ARM, Python 3.10 to 3.13. No Windows support.
Output formatsMarkdown, JSON, HTML, LaTeX, bounding boxesALTO, PageXML, hOCR, plain text, bounding boxes
BenchmarksOmniDocBench v1.5 93.12, olmOCR-Bench 79.8, OCRBench 880 (reported by the authors)Not published
Languages and scripts192 languages, handwriting and scene textRight-to-left, BiDi and top-to-bottom scripts, with community models on Zenodo
Known failureSpeed of 1.024 pages per second on an A100 even with W8A8. Thinking mode can emit up to 16,384 tokens, adding latency. Not for CPU-only deployment.Install can fail with a dependency conflict (issue #816). Learned line order is lost in segmentation and text output (issue #820). Usually needs training for good accuracy.

Sources

What each one is

Qianfan OCR

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

Qianfan OCR details

Kraken

OCR engine for historical documents and manuscripts, with trainable models for unusual scripts.

Kraken 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": ["kraken/kraken"],"file_url": "https://example.com/document.pdf"}

Questions

More comparisons

Qianfan OCR alternativesKraken alternativesBest OCR models in 2026