DeepSeek-OCR vs MinerU
Pick DeepSeek-OCR for cheap, self-hosted bulk OCR on a CUDA GPU under a plain MIT licence. Pick MinerU for mixed PDF and Office corpora, Chinese and English academic papers, or CPU-only batch work. The main split is scope: DeepSeek-OCR is a single OCR model, while MinerU is a full parsing toolkit with several tiers and many output formats.
DeepSeek-OCRdeepseek/deepseek-ocr | MinerUopendatalab/mineru | |
|---|---|---|
| Type | Open source | Open source |
| Price | 1 credit / page | 2 credits / page |
| Per 1,000 pages | $1.00 | $2.00 |
| Typical speed | 1.1 s / page | 2.0 s / page |
| Best for | Cheap bulk conversion | Academic PDFs |
| Inputs | PDF, images | PDF, images |
| Languages | 100+ languages | 80+ languages |
| Max pages per request | 200 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | No |
| 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 DeepSeek-OCR if
- The MIT licence allows commercial use with no extra conditions.
- You can set the resolution from 64 to 400 vision tokens per page, and the authors quote about 2,500 tokens per second on one A100-40G.
- It suits research on token-efficient OCR, and cheap bulk OCR where you only pay for compute.
Choose MinerU if
- It accepts PDF, images, DOCX, PPTX, XLSX, EPUB, OFD and HTML in one tool, which fits mixed corpora for RAG.
- The basic ONNX tier runs on CPU with about 2 GB RAM, so you can start without a GPU.
- It can output Markdown, HTML, LaTeX, DOCX, EPUB and JSON, and it has an independent olmOCR-Bench overall score of 75.2 (MinerU 2.5.4).
Key differences
| DeepSeek-OCR | MinerU | |
|---|---|---|
| Licence | MIT | MinerU Open Source License (Apache-2.0 based with extra conditions); the MinerU2.5 VLM weights are tagged AGPL-3.0 |
| Hardware | CUDA GPU; reference setup is CUDA 11.8 and PyTorch 2.6 in bfloat16 | Basic tier runs on CPU with 2 GB RAM; VLM tiers need 16 GB RAM and a GPU with 8 GB+ VRAM |
| Output formats | Markdown, plain text, HTML, bounding boxes | Markdown, HTML, LaTeX, DOCX, EPUB, JSON |
| Input types | Page images and PDFs | PDF, images, DOCX, PPTX, XLSX, EPUB, OFD and HTML |
| Throughput | About 2,500 tokens per second on one A100-40G (authors' figure) | 2.12 pages per second on one A100 with vLLM (VLM tier) |
| Benchmarks | Not published | olmOCR-Bench overall 75.2 (MinerU 2.5.4); ParseBench mean 45.9 (MinerU2.5) |
| Known failure | Repeated tokens and looping over table headers (GitHub issues #12 and #16); it corrects nonsense words instead of copying them | VLM tiers drop merged header rows in tables such as bank statement account lines (issue #5554) and can produce repeated content (issue #5728) |
Sources
- DeepSeek-OCR GitHub
- Visual Merit or Linguistic Crutch paper
- MinerU GitHub README
- MinerU2.5 model card
- olmOCR-Bench table
- Issue #5554 table header rows
What each one is
MinerU
Open-source PDF to Markdown pipeline popular for academic and Chinese documents.
MinerU detailsTry both on your own file
Run the file once with each model and compare the output.
PDF, PNG, JPG, WebP, HEIC. Up to 20 MB each, 10 files.No file?
Use one, fall back to the other
{"model": "deepseek/deepseek-ocr","fallbacks": ["opendatalab/mineru"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
DeepSeek-OCR vs DoclingDeepSeek-OCR vs Unlimited OCRJina OCR vs MinerUGoogle Document AI vs MinerUMinerU vs DoclingMinerU vs Qwen-VLDeepSeek-OCR vs TrOCRDeepSeek-OCR vs Qwen-VL
DeepSeek-OCR alternativesMinerU alternativesBest OCR models in 2026