Granite Docling vs InternVL3.5
Pick Granite Docling for English PDFs with tables, code and equations, especially inside a Docling pipeline on CPU or a small GPU. Pick InternVL3.5 for document question answering or when you want one model family you can scale from 1B to 241B. The main split is output: Granite Docling returns structured DocTags with bounding boxes, while InternVL3.5 returns plain text, Markdown or JSON with no layout by default.
Granite Doclingibm/granite-docling | InternVL3.5opengvlab/internvl | |
|---|---|---|
| Type | Open source | Vision model |
| Price | 1 credit / page | 2 credits / page |
| Per 1,000 pages | $1.00 | $2.00 |
| Typical speed | 1.5 s / page | 3.0 s / page |
| Best for | Small-model document conversion | Self-hosted document understanding |
| Inputs | PDF, images | images |
| Languages | English, experimental CJK and Arabic | Multilingual |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | No |
| Math | ✓ Yes | ✓ Yes |
| Bounding boxes | ✓ Yes | No |
| Confidence | No | No |
| Forms | No | No |
| Multilingual | No | ✓ Yes |
| 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
- Its 258M parameters keep conversion cheap on CPU or a small GPU.
- It outputs DocTags, Markdown, HTML, LaTeX and bounding boxes, so layout and coordinates are available.
- The model card reports table structure TEDS of 0.97 on FinTabNet, equation F1 of 0.968 and code recognition F1 of 0.988.
Choose InternVL3.5 if
- It comes in nine sizes from 1B to 241B-A28B, so you can trade accuracy for cost within one family.
- It suits document question answering and wider multimodal pipelines, not only page conversion.
- Models up to 30B run on a single A100, and the 20B-A4B and 30B-A3B MoE variants activate few parameters per token.
Key differences
| Granite Docling | InternVL3.5 | |
|---|---|---|
| Licence | Apache-2.0 | Apache-2.0, built on Apache-2.0 Qwen3 |
| Size | 258M parameters | 1B to 241B-A28B; the 8B variant is 8.5B |
| Output formats | DocTags, Markdown, HTML, LaTeX, bounding boxes | Plain text, Markdown, JSON; no bounding boxes or layout output by default |
| Published OCR benchmarks | Full-page OCR F1 0.84, OCRBench 500, table TEDS 0.97 (FinTabNet), all from the model card | Not published as text; the model card shows OCR scores only as chart images |
| Languages | English primary; Japanese, Arabic and Chinese are marked experimental | Not published |
| Hardware | Runs on CPU or a small GPU; T4-class GPUs without bfloat16 may output only exclamation marks unless run in float32 | Up to 30B on one A100, 38B on two A100s, 241B-A28B on eight A100s |
| Known failure | Can get stuck in repetitive loops on page headings and invoices; vLLM can fail with a wte attribute error unless you use the untied revision | Larger sizes are impractical on CPU-only setups; no word coordinates |
Sources
- Granite-Docling model card
- HF discussion #24 repetition
- Ollama issue #12355
- Hugging Face model card (8B)
- InternVL GitHub
What each one is
Granite Docling
Compact IBM vision model that converts pages to DocTags for the Docling pipeline.
Granite Docling detailsInternVL3.5
Shanghai AI Lab's open vision-language family (1B to 241B, August 2025), strong on OCR and document benchmarks.
InternVL3.5 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": ["opengvlab/internvl"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Granite Docling alternativesInternVL3.5 alternativesBest OCR models in 2026