docTR vs Qwen-VL
Pick docTR if you need word-level text with coordinates from invoices and receipts and want a small Python library you can train yourself. Pick Qwen-VL if you need layout-aware output such as HTML or Markdown, OCR in 32 languages, or one model that also answers questions about the page. Watch Qwen-VL for repetition loops on low-quality scans, which docTR's two-stage design does not report.
docTRmindee/doctr | Qwen-VLqwen/qwen-vl | |
|---|---|---|
| Type | Open source | Open source |
| Price | 1 credit / page | 2 credits / page |
| Per 1,000 pages | $1.00 | $2.00 |
| Typical speed | 0.9 s / page | 2.0 s / page |
| Best for | Word-level OCR with positions | Flexible extraction with prompts |
| Inputs | PDF, images | PDF, images |
| Languages | Mainly Latin scripts | 30+ languages |
| Max pages per request | 500 | 200 |
| Tables | No | ✓ Yes |
| Handwriting | No | ✓ Yes |
| Math | No | ✓ Yes |
| Bounding boxes | ✓ Yes | ✓ Yes |
| Confidence | ✓ Yes | No |
| Forms | No | 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 docTR if
- You want nested page, block, line and word JSON with coordinates from a library that returns words and lines only.
- You run Python on CPU or an NVIDIA GPU and want ready Docker images and a Hugging Face Spaces demo.
- You plan to train your own detection and recognition models, or add a key information extraction step for invoices and receipts.
Choose Qwen-VL if
- You need Markdown, HTML, JSON or bounding box output rather than words and lines.
- You need OCR across 32 languages, or OCR plus reasoning or visual question answering in one model.
- You work with photos and screenshots with text, and want sizes from 2B to 235B-A22B MoE under Apache-2.0.
Key differences
| docTR | Qwen-VL | |
|---|---|---|
| What it is | Two-stage text detection and recognition library from Mindee, built on PyTorch | General vision-language model family from Alibaba Qwen, latest Qwen3-VL |
| Licence | Apache-2.0 | Apache-2.0 (Qwen3-VL repository) |
| How you run it | Local Python library with Docker images; GPU images need CUDA 12.2 or newer on the host | Self-hosted open weights, or Alibaba Cloud hosted API; FP8 checkpoints need an H100-class GPU and CUDA 12+ |
| Price | Free, self-hosted (you pay for compute) | Free, self-hosted (you pay for compute); hosted Alibaba Cloud pricing varies by model |
| Output formats | JSON, plain text, bounding boxes. No table, formula or Markdown output | Plain text, HTML, bounding boxes, JSON, Markdown |
| Benchmark | Not published | Qwen 3 VL 8B scores 64.6 on olmOCR-bench in the Datalab table, below dedicated OCR models |
| Known failure | Cropping images leads to poorer results (issue #2007); split-crop merging misjoins readings across repeated characters (issue #2229) | Can loop repeating the same text, worse with greedy decoding (issue #1611); llama.cpp GGUF of Qwen2.5-VL-32B often loops on scanned documents |
Sources
- docTR GitHub README
- Issue #2007
- Issue #2229
- Qwen3-VL GitHub
- Issue #1611 infinite repetition
- Chandra README benchmark table
What each one is
docTR
Open-source OCR library from Mindee with separate text detection and recognition models.
docTR detailsQwen-VL
Alibaba's open-weight vision language models (Qwen2.5-VL, Qwen3-VL). Strong general OCR and grounding.
Qwen-VL 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": "mindee/doctr","fallbacks": ["qwen/qwen-vl"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
docTR alternativesQwen-VL alternativesBest OCR models in 2026