OCRRouter

Granite Docling vs GOT-OCR 2.0

Pick Granite Docling for English PDFs with tables, code and equations, especially inside a Docling pipeline on CPU or a small GPU. Pick GOT-OCR 2.0 for formulas, sheet music, chemical structures, region-specific OCR, or English and Chinese pages. The main split is scope: Granite Docling (258M) converts documents into structured output, while GOT-OCR 2.0 (580M) reads a wider range of content types.

Granite Doclingibm/granite-doclingGOT-OCR 2.0stepfun/got-ocr-2
TypeOpen sourceOpen source
Price1 credit / page1 credit / page
Per 1,000 pages$1.00$1.00
Typical speed1.5 s / page1.5 s / page
Best forSmall-model document conversionFormatted text, formulas and charts
InputsPDF, imagesimages
LanguagesEnglish, experimental CJK and ArabicChinese, English
Max pages per request500100
Tables✓ Yes✓ Yes
HandwritingNoNo
Math✓ Yes✓ Yes
Bounding boxes✓ Yes✓ Yes
ConfidenceNoNo
FormsNoNo
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 Granite Docling if

  • You want table structure from English PDFs: IBM reports TEDS 0.97 on FinTabNet at 150dpi.
  • You run Docling and want output as DocTags, Markdown, HTML, LaTeX or bounding boxes from a 258M-parameter model that is Apache-2.0 licensed.
  • You need low-cost conversion on CPU or a small GPU, and your pages are mostly English with code and equations (code F1 0.988, equation F1 0.968 per the model card).

Choose GOT-OCR 2.0 if

  • You need formulas, sheet music, chemical structures or charts, with output as LaTeX, TikZ, SMILES or kern.
  • You want to OCR one region of a page, guided by box coordinates or color, or run multi-page OCR with dynamic resolution.
  • Your documents are English or Chinese. Granite Docling marks Chinese as experimental.

Key differences

Granite DoclingGOT-OCR 2.0
Size258M parameters580M parameters
LicenceApache-2.0Code Apache-2.0; training data CC BY-NC 4.0
Output formatsDocTags, Markdown, HTML, LaTeX, bounding boxesPlain text, Markdown, LaTeX, TikZ, SMILES, kern
How you run itSelf-hosted, typically inside Docling. On GPUs without bfloat16 such as T4, run in float32 or it may output only exclamation marks. vLLM needs the untied revision.Self-hosted, available in Hugging Face Transformers with batched inference. README lists CUDA 11.8 and PyTorch 2.0.1; CPU users are pointed to a llama.cpp port.
Published benchmarksTable TEDS 0.97 (FinTabNet), full-page OCR F1 0.84, equation F1 0.968, OCRBench 500, all from the authorsNot published
LanguagesEnglish primary; Japanese, Arabic and Chinese experimentalEnglish and Chinese documents
Known failureRepetitive loops on page headings and invoices (HF discussion #24)Looping characters midway through recognition (issue #226); wrong text instead of empty output on images with no text (issue #235); format mode much worse than plain OCR mode (issue #317)

Sources

What each one is

Granite Docling

Compact IBM vision model that converts pages to DocTags for the Docling pipeline.

Granite Docling details

GOT-OCR 2.0

Unified OCR model that reads plain text, tables, formulas, sheet music and charts.

GOT-OCR 2.0 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": "ibm/granite-docling","fallbacks": ["stepfun/got-ocr-2"],"file_url": "https://example.com/document.pdf"}

Questions

More comparisons

Granite Docling alternativesGOT-OCR 2.0 alternativesBest OCR models in 2026