olmOCR vs Qwen-VL
Pick olmOCR for high-volume PDF to Markdown, such as academic papers with equations or pages that mix print and handwriting. Pick Qwen-VL when you want one general model for OCR plus reasoning or VQA, or for photos and screenshots with text. The main reason: olmOCR scores 82.4 on olmOCR-Bench, while Qwen3-VL 8B scores 64.6 in an independent table.
olmOCRallenai/olmocr | 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 | Bulk PDF to text for datasets | Flexible extraction with prompts |
| Inputs | PDF, images | PDF, images |
| Languages | Mainly English | 30+ languages |
| Max pages per request | 1,000 | 200 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | ✓ Yes | ✓ Yes |
| Math | ✓ Yes | ✓ 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 olmOCR if
- It scores 82.4 on olmOCR-Bench, above Marker (76.1), MinerU 2.5.4 (75.2) and Mistral OCR (72.0) in the same table.
- Self-hosting costs under $200 per million pages, per Ai2, and the weights, training data and code are all released.
- It handles equations, tables and handwriting, and strips headers and footers in natural reading order.
Choose Qwen-VL if
- It comes in sizes from 2B to 235B-A22B MoE, all Apache-2.0, so you can match the model to your hardware.
- It returns layout positions, bounding boxes and a Qwen HTML document format, not just plain text.
- It does OCR in 32 languages and handles low light, blur and tilted text, per Qwen.
Key differences
| olmOCR | Qwen-VL | |
|---|---|---|
| Model size | 7B | 2B, 4B, 8B, 32B dense; 30B-A3B and 235B-A22B MoE |
| Licence | Apache-2.0 (toolkit) | Apache-2.0 (Qwen3-VL repository) |
| Hardware | NVIDIA GPU with at least 12 GB VRAM and 30 GB free disk. Out of memory on a 12 GB GPU in issue #424. | FP8 checkpoints need an H100-class GPU and CUDA 12+. The 235B example uses 8-way tensor parallel on vLLM. |
| Price | Self-hosted: under $200 per million pages. Hosted olmOCR-2-7B on DeepInfra: $0.09 input / $0.19 output per million tokens. | Free when self-hosted, you pay for compute. Hosted pricing on Alibaba Cloud Model Studio: Not published. |
| Output formats | Markdown, JSON | Plain text, HTML, bounding boxes, JSON, Markdown |
| olmOCR-Bench score | 82.4 overall (v0.4.0), reported by the authors | 64.6 for Qwen 3 VL 8B, in an independent Datalab table |
| Known failure | Tends to get stuck on the last few pages (issue #433). Image bounding box coordinates are unreliable (issue #437). | Can loop repeating the same text, worse with greedy decoding and on low-quality images (issues #1611, #241). transformers 5.17 caused loops on 10-37% of OCR outputs versus 2-3% baseline. |
Sources
- olmOCR GitHub README
- Issue #424 OOM
- Issue #437 bbox
- Qwen3-VL GitHub
- Issue #1611 infinite repetition
- Chandra README benchmark table
What each one is
olmOCR
Open-source vision model from Ai2 (often searched as olm ocr) that turns PDFs into clean, reading-order text. Latest release: olmOCR 2.
olmOCR 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": "allenai/olmocr","fallbacks": ["qwen/qwen-vl"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
olmOCR alternativesQwen-VL alternativesBest OCR models in 2026