Qwen-VL vs Donut
Pick Qwen-VL if you want one Apache-2.0 model that reads photos, screenshots and documents and outputs Markdown, HTML, JSON or bounding boxes without training. Pick Donut if your documents are a narrow, fixed-layout type such as receipts or forms and you will fine-tune it to return JSON. The main split is that Qwen-VL works out of the box but can loop on low-quality scans, while Donut needs training before it is usable.
Qwen-VLqwen/qwen-vl | Donutnaver/donut | |
|---|---|---|
| Type | Open source | Open source |
| Price | 2 credits / page | 1 credit / page |
| Per 1,000 pages | $2.00 | $1.00 |
| Typical speed | 2.0 s / page | 1.0 s / page |
| Best for | Flexible extraction with prompts | Receipt and form fields after fine-tuning |
| Inputs | PDF, images | images |
| Languages | 30+ languages | English, Chinese, Japanese, Korean |
| Max pages per request | 200 | 100 |
| Tables | ✓ Yes | No |
| Handwriting | ✓ Yes | No |
| Math | ✓ Yes | 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 Qwen-VL if
- It runs without fine-tuning and outputs plain text, HTML, bounding boxes, JSON and Markdown.
- It does OCR in 32 languages and comes in sizes from 2B to 235B-A22B MoE, all under Apache-2.0, with a hosted API on Alibaba Cloud if you do not want to run GPUs.
- It handles OCR plus reasoning or VQA in one model, and copes with photos and screenshots with text, including low light, blur and tilted text per Qwen.
Choose Donut if
- It maps document images straight to structured JSON with no separate OCR step, which suits receipt and form field extraction once fine-tuned.
- It is MIT licensed and loads in Hugging Face Transformers, with no hosted service tied to it.
- A practitioner on Hacker News reports about 200 documents of fine-tuning can reach human-level accuracy on narrow data.
Key differences
| Qwen-VL | Donut | |
|---|---|---|
| Licence | Apache-2.0 | MIT |
| How you run it | Self-host the open weights, or use the hosted API on Alibaba Cloud | Self-host only |
| Ready to use | Yes, general model with OCR and document parsing | No, the base checkpoint is pretrain-only and must be fine-tuned |
| Output formats | Plain text, HTML, bounding boxes, JSON, Markdown | JSON, plain text |
| Published benchmark | Qwen 3 VL 8B scores 64.6 on olmOCR-bench in the Datalab table, below dedicated OCR models | Not published |
| Known failure | Can loop and repeat the same text, mostly on low-quality images and worse with greedy decoding; GGUF builds of Qwen2.5-VL-32B often loop on scanned documents | Misses tiny text in large images because of input resolution limits |
| Setup notes | FP8 checkpoints need an H100-class GPU and CUDA 12+ | Transformers v5 dropped image-to-text pipeline support, so load the model directly or use v4 |
Sources
- Qwen3-VL GitHub
- Issue #1611 infinite repetition
- Chandra README benchmark table
- donut-base model card
- Donut ECCV paper
- Hacker News thread
What each one is
Qwen-VL
Alibaba's open-weight vision language models (Qwen2.5-VL, Qwen3-VL). Strong general OCR and grounding.
Qwen-VL 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.
Use one, fall back to the other
{"model": "qwen/qwen-vl","fallbacks": ["naver/donut"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Qwen-VL alternativesDonut alternativesBest OCR models in 2026