OCRRouter

dots.ocr vs Qwen-VL

Pick dots.ocr for faithful document parsing: multilingual PDFs to Markdown with layout, tables as HTML and formulas as LaTeX, from a 3B self-hosted model. Pick Qwen-VL when you need one general model for OCR plus reasoning or VQA, or a hosted API. For high-volume transcription, the main reason to lean to dots.ocr is that Qwen-VL can loop and repeat text, and its 8B model scores 64.6 on olmOCR-bench against 79.1 for the original dots.ocr.

dots.ocrrednote/dots-ocrQwen-VLqwen/qwen-vl
TypeOpen sourceOpen source
Price2 credits / page2 credits / page
Per 1,000 pages$2.00$2.00
Typical speed1.5 s / page2.0 s / page
Best forMultilingual document parsingFlexible extraction with prompts
InputsPDF, imagesPDF, images
Languages100+ languages30+ languages
Max pages per request500200
Tables✓ Yes✓ Yes
HandwritingNo✓ Yes
Math✓ Yes✓ Yes
Bounding boxes✓ Yes✓ Yes
ConfidenceNoNo
FormsNoNo
Multilingual✓ Yes✓ Yes
On OCRRouterAvailableAvailable

Prices are what you pay on OCRRouter. Speed is the provider's typical figure, not a benchmark we ran.

Which to choose

Choose dots.ocr if

  • It scores 79.1 on olmOCR-bench (83.9 for dots.mocr, both reported by the authors), against 64.6 for Qwen 3 VL 8B in an independent table.
  • One 3B model returns layout boxes, categories and text, with tables as HTML and formulas as LaTeX, for academic and business documents with mixed tables and text.
  • The dots.mocr-svg variant converts charts and diagrams to SVG, which Qwen-VL does not list.

Choose Qwen-VL if

  • It also serves as a general model for OCR plus reasoning or VQA, and handles photos and screenshots with text.
  • You can use open weights self-hosted or the hosted API on Alibaba Cloud, and sizes run from 2B to 235B-A22B MoE, all Apache-2.0.
  • It reads OCR in 32 languages and handles low light, blur and tilted text per Qwen.

Key differences

dots.ocrQwen-VL
LicenceMIT (the repository also includes a separate dots.ocr license agreement file)Apache-2.0 (Qwen3-VL repository)
How you run itSelf-host only, via vLLM or Hugging Face Transformers. A single CUDA GPU with vLLM is the documented path.Self-host or hosted API on Alibaba Cloud. FP8 checkpoints need an H100-class GPU and CUDA 12+.
Model size3B (1.7B LLM backbone)2B, 4B, 8B, 32B dense; 30B-A3B and 235B-A22B MoE
PriceFree, self-hosted (you pay for compute)Free, self-hosted (you pay for compute). Hosted pricing varies by model: Not published here.
Output formatsJSON, Markdown, HTML, LaTeX, bounding boxes, SVGPlain text, HTML, bounding boxes, JSON, Markdown
olmOCR-bench79.1 (dots.ocr), 83.9 (dots.mocr), reported by the authors64.6 (Qwen 3 VL 8B), from an independent table
Known failureRuns of ellipses or underscores can make output repeat endlessly. A table above 1000 grid cells raises an error and drops the whole page in Docling.Can loop repeating the same text, worse with greedy decoding and on low-quality images. One report saw loops on 10-37% of OCR outputs with transformers 5.17, versus 2-3% at baseline.

Sources

What each one is

dots.ocr

Single vision model from rednote that does layout detection and content recognition together.

dots.ocr details

Qwen-VL

Alibaba's open-weight vision language models (Qwen2.5-VL, Qwen3-VL). Strong general OCR and grounding.

Qwen-VL details

Try 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": "rednote/dots-ocr","fallbacks": ["qwen/qwen-vl"],"file_url": "https://example.com/document.pdf"}

Questions

More comparisons

dots.ocr alternativesQwen-VL alternativesBest OCR models in 2026