Surya vs Qwen-VL
Pick Surya for local, high-volume transcription of printed documents where you need layout labels, reading order and bounding boxes. It is 689M parameters, scores 83.7% on olmOCR-bench, and runs on a laptop CPU or Apple Silicon. Pick Qwen-VL when the job mixes OCR with reasoning or VQA, or covers photos and screenshots with text, and accept that you must guard against repetition loops.
Suryadatalab/surya | Qwen-VLqwen/qwen-vl | |
|---|---|---|
| Type | Open source | Open source |
| Price | 2 credits / page | 2 credits / page |
| Per 1,000 pages | $2.00 | $2.00 |
| Typical speed | 1.5 s / page | 2.0 s / page |
| Best for | Layout and line-level OCR | Flexible extraction with prompts |
| Inputs | PDF, images | PDF, images |
| Languages | 90+ languages | 30+ languages |
| Max pages per request | 500 | 200 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | ✓ Yes |
| Math | No | ✓ Yes |
| Bounding boxes | ✓ Yes | ✓ Yes |
| Confidence | No | No |
| Forms | No | No |
| Multilingual | ✓ Yes | ✓ 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 Surya if
- It scores 83.7% on olmOCR-bench with 689M parameters, while Qwen 3 VL 8B scores 64.6 in an independent table.
- It runs on CPU and Apple Silicon through llama.cpp, with peak memory of about 4.3-5.9 GB on an M4 Max.
- One model returns layout labels, reading order, OCR, HTML tables and KaTeX math with per-block bounding boxes in 90+ languages.
Choose Qwen-VL if
- It is a general vision-language model, so one deployment covers OCR, reasoning and VQA on photos and screenshots.
- It comes in sizes from 2B to 235B-A22B MoE under Apache-2.0, and Alibaba Cloud also serves it as a hosted API.
- It handles low light, blur and tilted text per Qwen, and outputs layout positions and a Qwen HTML document format.
Key differences
| Surya | Qwen-VL | |
|---|---|---|
| Licence | Code Apache-2.0; weights under a modified AI Pubs Open Rail-M. Free for research, personal use and companies under $5M funding or revenue. | Apache-2.0 for the Qwen3-VL repository, with no revenue cap listed. |
| How you run it | Self-host only. A lightweight client talks to a separately running vLLM or llama.cpp server. | Self-host the open weights, or use the hosted API on Alibaba Cloud. FP8 checkpoints need an H100-class GPU and CUDA 12+. |
| Model size | 689M parameters. | 2B, 4B, 8B, 32B dense; 30B-A3B and 235B-A22B MoE. |
| olmOCR-bench | 83.7% (Surya 2.5, reported by the authors). | 64.6 for Qwen 3 VL 8B (independent, in the Chandra table). |
| Languages | 90+ languages. The 91-language pass rate is 74.9%, and only 21 languages reach 90%. | OCR in 32 languages, up from 10 in the previous generation. |
| Known failure | Surya 2.5 table output lacks row/column geometry, and multilingual tables score 58.1. Handwriting, complex tables and math are better sent to the hosted Datalab API. | Can loop and repeat the same text, worse with greedy decoding and on low-quality images. One report saw loops on 10-37% of OCR outputs versus 2-3% baseline, and a Qwen2.5-VL-32B GGUF often got stuck on scanned documents. |
| Output formats | JSON, HTML, LaTeX, bounding boxes, plain text. | Plain text, HTML, bounding boxes, JSON, Markdown. |
Sources
- Surya GitHub README
- Datalab pricing and license FAQ
- Qwen3-VL GitHub
- Issue #1611 infinite repetition
- olmo-eval issue #441
- Chandra README benchmark table
What each one is
Surya
Open-source OCR toolkit with line detection, layout analysis, reading order and tables.
Surya 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": "datalab/surya","fallbacks": ["qwen/qwen-vl"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Surya alternativesQwen-VL alternativesBest OCR models in 2026