Marker vs Qwen-VL
Pick Marker for self-hosted PDF to Markdown on books and papers with equations, or for RAG chunking, if your company's licence situation allows it. Pick Qwen-VL when you want one Apache-2.0 model that also handles photos, screenshots and reasoning over the text. The main split is purpose: Marker is a dedicated document converter that scores 83.7 on olmOCR-bench (Datalab's own figure), while Qwen 3 VL 8B scores 64.6 and can loop on poor scans.
Markerdatalab/marker | Qwen-VLqwen/qwen-vl | |
|---|---|---|
| Type | OCR API | Open source |
| Price | 14 credits / page | 2 credits / page |
| Per 1,000 pages | $14.00 | $2.00 |
| Typical speed | 2.5 s / page | 2.0 s / page |
| Best for | Books and research papers | Flexible extraction with prompts |
| Inputs | PDF, images, DOCX | PDF, images |
| Languages | 90+ languages | 30+ languages |
| Max pages per request | 1,000 | 200 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | ✓ Yes |
| Math | ✓ Yes | ✓ 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 Marker if
- You need PDF to Markdown, HTML, JSON block trees or RAG chunks out of the box, including LaTeX for papers and books.
- You have no GPU: fast mode runs on CPU with about 18 MB of ONNX weights at 23.6 pages per second, though it drops to 66.6 on olmOCR-bench.
- You want a hosted option with about 5,000 free pages per month from Datalab, or local use on GPU, CPU or Apple Silicon.
Choose Qwen-VL if
- You need a permissive licence: all sizes are Apache-2.0, with no revenue cap on commercial use.
- You also need reasoning or VQA on the same input, or you process photos and screenshots with text, including blur, low light and tilted text.
- You need OCR in 32 languages, bounding boxes for layout positions, or a size from 2B up to 235B-A22B MoE.
Key differences
| Marker | Qwen-VL | |
|---|---|---|
| Licence | Code Apache-2.0. Weights under a modified AI Pubs Open Rail-M licence: free for research, personal use and startups under $5M. | Apache-2.0 (Qwen3-VL repository). |
| Commercial use | Companies above $5M need a paid licence for the weights. | No such restriction listed. |
| What it is | PDF to Markdown converter built on the Surya OCR models. | General vision-language model family (latest Qwen3-VL) with OCR and document parsing. |
| Output formats | Markdown, HTML, JSON, chunks, LaTeX, images. | Plain text, HTML, bounding boxes, JSON, Markdown. |
| Hardware | GPU, CPU or MPS. vllm mode needs an NVIDIA GPU. Fast mode runs on CPU. | FP8 checkpoints need an H100-class GPU and CUDA 12+. The 235B example uses 8-way tensor parallel on vLLM. |
| olmOCR-bench | 83.7 in vllm mode and 66.6 in CPU fast mode (reported by Datalab). | 64.6 for Qwen 3 VL 8B (Datalab table). |
| Known failure | Not published for output quality. All README benchmark numbers are Datalab's own. | Can loop and repeat the same text, worse with greedy decoding and on low-quality images. Reported on scanned documents with a Qwen2.5-VL-32B GGUF in llama.cpp. |
Sources
- Marker GitHub README
- Datalab pricing
- Qwen3-VL GitHub
- Issue #1611 infinite repetition
- llama.cpp issue #14663
- Chandra README benchmark table
What each one is
Marker
Hosted version of the open-source Marker and Surya models. Strong on books and papers.
Marker 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/marker","fallbacks": ["qwen/qwen-vl"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Marker alternativesQwen-VL alternativesBest OCR models in 2026