Qwen-VL vs GOT-OCR 2.0
Pick Qwen-VL when you need OCR together with reasoning or question answering on photos, screenshots and mixed content, and you can add a check for repeated-text loops. Pick GOT-OCR 2.0 when the job is formulas, sheet music, chemical structures or reading a specific region of a page on a small 580M model. The main split is scope: Qwen-VL is a general vision-language family from 2B to 235B, GOT-OCR 2.0 is one small OCR-only model.
Qwen-VLqwen/qwen-vl | GOT-OCR 2.0stepfun/got-ocr-2 | |
|---|---|---|
| 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 | 1.5 s / page |
| Best for | Flexible extraction with prompts | Formatted text, formulas and charts |
| Inputs | PDF, images | images |
| Languages | 30+ languages | Chinese, English |
| Max pages per request | 200 | 100 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | ✓ Yes | No |
| Math | ✓ Yes | ✓ Yes |
| Bounding boxes | ✓ Yes | ✓ Yes |
| Confidence | No | No |
| Forms | No | No |
| 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
- You want one model for OCR plus reasoning or VQA, with outputs as plain text, HTML, bounding boxes, JSON or Markdown.
- You need OCR in 32 languages, and a size from 2B up to 235B-A22B MoE, all under Apache-2.0.
- You want the option of a hosted API on Alibaba Cloud as well as self-hosting.
Choose GOT-OCR 2.0 if
- You read formulas, sheet music, chemical structures, charts or geometry and want LaTeX, TikZ, SMILES or kern output.
- You need region OCR guided by box coordinates or color, plus multi-page OCR and dynamic resolution.
- You want a 580M model that is merged into Hugging Face Transformers with batched inference.
Key differences
| Qwen-VL | GOT-OCR 2.0 | |
|---|---|---|
| Licence | Apache-2.0 (Qwen3-VL repository) | Code Apache-2.0; data CC BY-NC 4.0 |
| How you run it | Self-host the open weights, or use the hosted API on Alibaba Cloud | Self-host only |
| Model size | 2B, 4B, 8B, 32B dense; 30B-A3B and 235B-A22B MoE | 580M |
| Output formats | Plain text, HTML, bounding boxes, JSON, Markdown | Plain text, Markdown, LaTeX, TikZ, SMILES, kern |
| Benchmark | Qwen 3 VL 8B scores 64.6 on olmOCR-bench (independent, Datalab table), below dedicated OCR models | Not published |
| Hardware | FP8 checkpoints need an H100-class GPU and CUDA 12+ | CUDA 11.8 and PyTorch 2.0.1 per README; CPU users are pointed to a llama.cpp port |
| Known failure | Can loop and repeat the same text, worse with greedy decoding and on low-quality scans; Qwen2.5-VL-32B GGUF often loops on scanned documents | Can loop characters midway through recognition, returns wrong text on images with no text, and format mode gives much worse results than plain OCR mode |
Sources
- Qwen3-VL GitHub
- Issue #1611 infinite repetition
- Chandra README benchmark table
- llama.cpp issue #14663
- GOT-OCR2.0 GitHub
- GOT-OCR 2.0 paper
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 detailsGOT-OCR 2.0
Unified OCR model that reads plain text, tables, formulas, sheet music and charts.
GOT-OCR 2.0 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": ["stepfun/got-ocr-2"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Qwen-VL alternativesGOT-OCR 2.0 alternativesBest OCR models in 2026