Granite Docling vs SmolDocling
Pick Granite Docling for any real conversion job. It is the successor to SmolDocling, scores higher on every reported task (table TEDS 0.97 vs 0.82), and is the model the authors point users to for updates and support. SmolDocling is only worth running for experiments with a tiny document model on a low budget.
Granite Doclingibm/granite-docling | SmolDoclinghuggingface/smoldocling | |
|---|---|---|
| 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 | Very cheap page conversion |
| Inputs | PDF, images | PDF, images |
| Languages | English, experimental CJK and Arabic | Mainly English |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | No |
| Math | ✓ Yes | ✓ Yes |
| Bounding boxes | ✓ Yes | ✓ Yes |
| Confidence | No | No |
| Forms | No | No |
| 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
- Table structure is far better: TEDS 0.97 on FinTabNet at 150dpi, against 0.82 for SmolDocling.
- It reports equation F1 0.968 and code recognition F1 0.988, so English PDFs with formulas and code are covered.
- It is the maintained release, while SmolDocling is a preview that the authors say is superseded.
Choose SmolDocling if
- It has a published speed figure of about 0.35 seconds per page on an A100 with vLLM.
- At 256M parameters it runs on modest hardware, and so does Granite Docling at 258M, so you give up little on size.
- It handles text, tables, code, formulas and charts in one model with location tags, which suits quick experiments with a small document model.
Key differences
| Granite Docling | SmolDocling | |
|---|---|---|
| Status | Current release from IBM | Preview release, superseded by Granite Docling |
| Parameters | 258M | 256M |
| Licence | Apache-2.0 | Apache-2.0 per model summary; metadata tag says CDLA-Permissive-2.0 |
| Table TEDS (FinTabNet, 150dpi) | 0.97 | 0.82 |
| Full-page OCR F1 / OCRBench | 0.84 / 500 | 0.80 / 338 |
| Languages | English primary; Japanese, Arabic and Chinese are experimental | English only |
| Known failure | Repetitive loops on page headings and invoices; exclamation-mark output on GPUs without bfloat16 such as T4 unless run in float32 | Endless token repetition and missing closing tags; output loops on some images make generation very slow |
Sources
- Granite-Docling model card
- IBM announcement
- HF discussion #24 repetition
- SmolDocling model card
- SmolDocling paper
- Docling issue #1515
What each one is
Granite Docling
Compact IBM vision model that converts pages to DocTags for the Docling pipeline.
Granite Docling detailsSmolDocling
Tiny 256M vision model from Hugging Face and IBM that outputs DocTags. Predecessor of Granite Docling.
SmolDocling detailsTry 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": ["huggingface/smoldocling"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Granite Docling vs MonkeyOCRQwen-VL vs SmolDoclingAmazon Textract vs Granite DoclingChandra vs Granite DoclingGranite Docling vs MiniCPM-V 4.5Azure Document Intelligence vs Granite DoclingGranite Docling vs PixtraldocTR vs SmolDocling
Granite Docling alternativesSmolDocling alternativesBest OCR models in 2026