Donut vs InternVL3.5
Pick Donut if you have a narrow, fixed-layout document type such as receipts or forms and can fine-tune a small model to output JSON. Pick InternVL3.5 if you want a general model for document question answering that works without training and comes in sizes from 1B to 241B-A28B. Donut is the cheaper path once trained, but the base checkpoint is useless until you fine-tune it.
Donutnaver/donut | 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 | 1.0 s / page | 3.0 s / page |
| Best for | Receipt and form fields after fine-tuning | Self-hosted document understanding |
| Inputs | images | images |
| Languages | English, Chinese, Japanese, Korean | Multilingual |
| Max pages per request | 100 | 500 |
| Tables | No | ✓ Yes |
| Handwriting | No | No |
| Math | No | ✓ Yes |
| Bounding boxes | No | No |
| Confidence | No | No |
| Forms | ✓ Yes | 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 Donut if
- It maps a document image straight to structured JSON with no separate OCR step, which suits receipt and form field extraction after fine-tuning.
- A practitioner on Hacker News reports that about 200 documents of fine-tuning can reach human-level accuracy on narrow data.
- It is MIT licensed and loads from Hugging Face Transformers, so you can adapt it to your own fixed-layout documents.
Choose InternVL3.5 if
- It comes in nine sizes from 1B to 241B, so you can trade accuracy for cost, and the MoE variants 20B-A4B and 30B-A3B activate few parameters per token.
- Models up to 30B run on a single A100, so a mid-size setup needs one GPU.
- It outputs plain text, Markdown or JSON, and it suits document question answering without a training step.
Key differences
| Donut | InternVL3.5 | |
|---|---|---|
| Licence | MIT | Apache-2.0 |
| Price | Free, self-hosted (you pay for compute) | Free, self-hosted (you pay for compute) |
| Out-of-the-box use | No. The base checkpoint is pretrain-only and must be fine-tuned. | General vision language model, usable for document question answering as released. |
| Output formats | JSON, plain text | Plain text, Markdown, JSON |
| Model sizes and hardware | One base model. Hardware needs: Not published. | 1B to 241B-A28B. Up to 30B on one A100, 38B on two A100s, 241B-A28B on eight A100s. |
| Layout and coordinates | Not published | No bounding boxes or layout output by default. |
| Known failure | Misses tiny text in large images because of input resolution limits. One HN user found results on personal documents did not match the promised scores. | The model card gives OCR scores only as chart images, not text values. |
| Library support | Transformers v5 dropped image-to-text pipeline support, so load the model directly or stay on v4. | Runs with Transformers, LMDeploy or vLLM. |
Sources
What each one is
Donut
OCR-free document understanding model. Fine-tuned for receipts, forms and document Q&A.
Donut 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": "naver/donut","fallbacks": ["opengvlab/internvl"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Donut alternativesInternVL3.5 alternativesBest OCR models in 2026