Donut vs Phi-4-multimodal
Pick Phi-4-multimodal if you want to ask questions about English documents without training anything. It reports DocVQA 93.2 and OCRBench 84.4 out of the box. Pick Donut if you have a narrow, fixed-layout document type such as receipts or forms and can fine-tune on your own labelled samples.
Donutnaver/donut | Phi-4-multimodalmicrosoft/phi-4-multimodal | |
|---|---|---|
| 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 | 2.5 s / page |
| Best for | Receipt and form fields after fine-tuning | Small model for OCR and questions |
| Inputs | images | images |
| Languages | English, Chinese, Japanese, Korean | English for vision |
| 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 | 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 Donut if
- You can fine-tune it on a narrow, fixed-layout set such as receipts or forms, and a practitioner on HN reports about 200 documents can reach human-level accuracy on such data.
- It maps the document image straight to JSON with no separate OCR step, which suits field extraction and document classification.
- It is MIT licensed and loads in Hugging Face Transformers, so you own the whole pipeline and pay only for compute.
Choose Phi-4-multimodal if
- It works without fine-tuning: Microsoft reports DocVQA 93.2, OCRBench 84.4, ChartQA 81.4 and InfoVQA 72.7 at 5.6B parameters.
- You can self-host it or call it through Azure AI Foundry, so you can start on a managed endpoint and move to your own GPUs later.
- It outputs plain text, Markdown or JSON, has a 128K token context, and also takes audio in the same model for mixed audio and image pipelines.
Key differences
| Donut | Phi-4-multimodal | |
|---|---|---|
| Ready to use | No. The base checkpoint is pretrain-only and must be fine-tuned first. | Yes. It is an instruct model with published document benchmark scores. |
| Deployment | Self-host only. | Self-host or Azure AI Foundry. |
| Licence and cost | MIT. Free, you pay for compute. | MIT. Free self-hosted, you pay for compute. |
| Output formats | JSON, plain text. | Plain text, Markdown, JSON. |
| Published benchmarks | Not published | DocVQA 93.2, OCRBench 84.4, ChartQA 81.4, InfoVQA 72.7 (authors' numbers). |
| Languages | Not published | Vision is English only. Performance drops in other languages. |
| 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 warns of repetitive or inconsistent output in long chats. |
| Hardware and software | Transformers v5 dropped image-to-text pipeline support, so load the model directly or use v4. | Uses flash attention on A100, A6000 and H100. V100 and older GPUs need eager attention. |
Sources
What each one is
Donut
OCR-free document understanding model. Fine-tuned for receipts, forms and document Q&A.
Donut detailsPhi-4-multimodal
Microsoft's 5.6B text, vision and audio model (February 2025). Scores 84.4 on OCRBench per its model card.
Phi-4-multimodal 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": "naver/donut","fallbacks": ["microsoft/phi-4-multimodal"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Donut vs Google Cloud Vision OCRDonut vs DeepSeek-VL2GOT-OCR 2.0 vs Phi-4-multimodalNougat vs Phi-4-multimodalTrOCR vs DonutAmazon Textract vs Phi-4-multimodalReducto Parse vs Phi-4-multimodalGranite Docling vs Donut
Donut alternativesPhi-4-multimodal alternativesBest OCR models in 2026