Granite Docling vs Donut
Pick Granite Docling for converting whole English PDFs with tables, code and equations into Markdown, HTML or DocTags, because it works out of the box at 258M parameters. Pick Donut only if you have a narrow, fixed-layout document type such as receipts or forms and can fine-tune it, because the base checkpoint is pretrain-only and does not do page conversion.
Granite Doclingibm/granite-docling | Donutnaver/donut | |
|---|---|---|
| Type | Open source | Open source |
| Price | 1 credit / page | 1 credit / page |
| Per 1,000 pages | $1.00 | $1.00 |
| Typical speed | 1.5 s / page | 1.0 s / page |
| Best for | Small-model document conversion | Receipt and form fields after fine-tuning |
| Inputs | PDF, images | images |
| Languages | English, experimental CJK and Arabic | English, Chinese, Japanese, Korean |
| Max pages per request | 500 | 100 |
| Tables | ✓ Yes | No |
| Handwriting | No | No |
| Math | ✓ Yes | No |
| Bounding boxes | ✓ Yes | No |
| Confidence | No | No |
| Forms | No | ✓ Yes |
| Multilingual | No | No |
| On OCRRouter | Available | Available |
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 need full-page conversion to Markdown, HTML, LaTeX or DocTags with bounding boxes, which is what it outputs.
- Your documents are English PDFs with tables, code and equations: model card scores are 0.97 table structure TEDS on FinTabNet and 0.968 equation F1.
- You want cheap self-hosting on CPU or a small GPU, with 258M parameters and an Apache-2.0 licence, and you already use Docling.
Choose Donut if
- You have a fixed-layout document type such as receipts or forms and want JSON field extraction after fine-tuning.
- You want document classification as well as parsing, since the model is meant to be fine-tuned for parsing, classification or VQA.
- You want an MIT licence and no external OCR step: it maps the document image straight to structured output and loads in Hugging Face Transformers.
Key differences
| Granite Docling | Donut | |
|---|---|---|
| Ready to use | Converts documents as published, no training step described | Base checkpoint is pretrain-only and must be fine-tuned first |
| Output formats | DocTags, Markdown, HTML, LaTeX, bounding boxes | JSON, plain text |
| Licence | Apache-2.0 | MIT |
| Published benchmarks | Table TEDS 0.97 (FinTabNet), full-page OCR F1 0.84, equation F1 0.968, OCRBench 500 (authors' figures) | Not published |
| Languages | English primary; Japanese, Arabic and Chinese marked experimental | Not published |
| Main failure mode | Repetitive loops on page headings and invoices; exclamation-mark output on GPUs without bfloat16 such as T4 unless run in float32 | Misses tiny text in large images because of input resolution limits |
| Setup friction | vLLM may need the untied revision; Ollama on CPU reports FP16 overflow hallucinations | Transformers v5 dropped image-to-text pipeline support, so load the model directly or use v4 |
Sources
- Granite-Docling model card
- HF discussion #24 repetition
- Ollama issue #12355
- donut-base model card
- Donut ECCV paper
What each one is
Granite Docling
Compact IBM vision model that converts pages to DocTags for the Docling pipeline.
Granite Docling detailsDonut
OCR-free document understanding model. Fine-tuned for receipts, forms and document Q&A.
Donut detailsTry both on your own file
Run the file once with each model and compare the output.
Use one, fall back to the other
{"model": "ibm/granite-docling","fallbacks": ["naver/donut"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Granite Docling alternativesDonut alternativesBest OCR models in 2026