Surya vs Donut
Pick Surya if you need OCR, layout detection, reading order and bounding boxes for 90+ languages that works without training, including on a laptop. Pick Donut only if you have a narrow, fixed-layout document type such as receipts or forms and can fine-tune the model on your own data. The main difference is that Surya is a ready OCR toolkit, while Donut is a base model that must be trained before use.
Suryadatalab/surya | Donutnaver/donut | |
|---|---|---|
| Type | Open source | Open source |
| Price | 2 credits / page | 1 credit / page |
| Per 1,000 pages | $2.00 | $1.00 |
| Typical speed | 1.5 s / page | 1.0 s / page |
| Best for | Layout and line-level OCR | Receipt and form fields after fine-tuning |
| Inputs | PDF, images | images |
| Languages | 90+ languages | English, Chinese, Japanese, Korean |
| Max pages per request | 500 | 100 |
| Tables | ✓ Yes | No |
| Handwriting | No | No |
| Math | No | No |
| Bounding boxes | ✓ Yes | No |
| Confidence | No | No |
| Forms | No | ✓ Yes |
| Multilingual | ✓ Yes | 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 Surya if
- It works without training: one model returns layout labels, reading order, OCR, HTML tables and KaTeX math with per-block bounding boxes.
- It scores 83.7% on olmOCR-bench (reported by the authors) with 689M parameters, and runs at 4.52 pages/s at 192 DPI on one H100 with vLLM.
- It runs on CPU and Apple Silicon through llama.cpp, with peak memory of about 4-6 GB on an M4 Max.
Choose Donut if
- It needs no separate OCR step: it maps a document image straight to structured JSON, which suits receipt and form field extraction after fine-tuning.
- It is MIT licensed, with no revenue or funding cap, and loads in Hugging Face Transformers.
- A Hacker News practitioner reports that about 200 documents of fine-tuning can reach human-level accuracy on narrow data.
Key differences
| Surya | Donut | |
|---|---|---|
| Licence | Code Apache-2.0; weights under a modified AI Pubs Open Rail-M, free for research, personal use and companies under $5M funding/revenue | MIT |
| Ready to use | Yes, runs as a client against a vLLM or llama.cpp server | No, the base checkpoint is pretrain-only and must be fine-tuned first |
| Output | JSON, HTML, LaTeX, bounding boxes, plain text | JSON, plain text |
| Hardware | NVIDIA GPU via vLLM, or CPU/Apple Silicon via llama.cpp; peak memory about 4.3-5.9 GB on an M4 Max | Not published |
| Benchmarks | olmOCR-bench 83.7%, OmniParseBench 85.9, 91-language pass rate 74.9% (all by the authors) | Not published |
| Large pages and small text | Not published | Misses tiny text in large images because of input resolution limits |
| Known failure | Multilingual tables score 58.1, and Surya 2.5 table output lacks row/column geometry; not for handwriting or photos of natural scenes | A Hacker News user found results on personal documents did not match the promised scores; Transformers v5 dropped image-to-text pipeline support, so you need direct loading or v4 |
Sources
- Surya GitHub README
- Datalab pricing and license FAQ
- donut-base model card
- Donut ECCV paper
- Hacker News thread
What each one is
Surya
Open-source OCR toolkit with line detection, layout analysis, reading order and tables.
Surya 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.
PDF, PNG, JPG, WebP, HEIC. Up to 20 MB each, 10 files.No file?
Use one, fall back to the other
{"model": "datalab/surya","fallbacks": ["naver/donut"],"file_url": "https://example.com/document.pdf"}