TrOCR vs Qwen-VL
Pick TrOCR when you already have cropped single-line images, mostly English handwriting or receipt lines, and want a small recognizer you can fine-tune. Pick Qwen-VL when the input is a full page, a photo or a screenshot, or when you need layout, JSON or Markdown output in one of 32 languages. The main split is scope: TrOCR reads one text line, Qwen-VL reads whole documents but can loop on poor scans.
TrOCRmicrosoft/trocr | 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 | 1.2 s / page | 2.0 s / page |
| Best for | Handwritten text lines | Flexible extraction with prompts |
| Inputs | images | PDF, images |
| Languages | English | 30+ languages |
| Max pages per request | 100 | 200 |
| Tables | No | ✓ Yes |
| Handwriting | ✓ Yes | ✓ Yes |
| Math | No | ✓ Yes |
| Bounding boxes | No | ✓ Yes |
| Confidence | No | 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 TrOCR if
- It reports 2.89 character error rate on the IAM handwriting set with the large model, per the paper.
- It ships separate checkpoints fine-tuned for handwritten (IAM), printed (SROIE) and scene text, so you can start from the one that matches your lines.
- It is MIT licensed and loads in a few lines with Hugging Face transformers, which makes it easy to fine-tune on your own line images.
Choose Qwen-VL if
- It reads full pages, photos and screenshots, and handles low light, blur and tilted text per Qwen.
- It outputs plain text, HTML, bounding boxes, JSON and Markdown, so you get layout positions and not just text.
- It covers OCR in 32 languages and comes in sizes from 2B to 235B-A22B MoE, all Apache-2.0, with a hosted API on Alibaba Cloud as an option.
Key differences
| TrOCR | Qwen-VL | |
|---|---|---|
| Input scope | Single text-line images only, per the model card. You need a separate line detector. | General vision-language model for OCR and document parsing, with layout positions in the output. |
| Output formats | Plain text only. No layout, table or page-level output. | Plain text, HTML, bounding boxes, JSON and Markdown. |
| Languages | Released checkpoints are English-focused. Other languages need fine-tuning. | OCR in 32 languages, up from 10 in the previous generation. |
| Licence | MIT (microsoft/unilm repository). | Apache-2.0 (Qwen3-VL repository). |
| How you run it | Locally through Hugging Face transformers. Free, you pay for compute. | Self-host the open weights (FP8 checkpoints need an H100-class GPU and CUDA 12+) or use Alibaba Cloud's hosted API. Hosted pricing: Not published. |
| Model size | Base 334M, large 558M parameters. | 2B, 4B, 8B, 32B dense; 30B-A3B and 235B-A22B MoE. |
| Known failure | Slow on CPU versus CRNN recognizers, even for one text line. | Can loop and repeat the same text, worse with greedy decoding and on low-quality scans. One run saw loops on 10-37% of OCR outputs versus 2-3% baseline. GGUF builds often get stuck on scanned documents. |
Sources
- TrOCR paper (arXiv 2109.10282)
- trocr-large-handwritten model card
- Qwen3-VL GitHub
- Issue #1611 infinite repetition
- olmo-eval issue #441
- llama.cpp issue #14663
What each one is
TrOCR
Transformer text recognition model. Reads single text lines, including handwriting.
TrOCR 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": "microsoft/trocr","fallbacks": ["qwen/qwen-vl"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
TrOCR alternativesQwen-VL alternativesBest OCR models in 2026