OCRRouter

DeepSeek-OCR vs DeepSeek-VL2

Pick DeepSeek-OCR for cheap self-hosted bulk transcription of pages to Markdown, HTML or bounding boxes under a plain MIT licence. Pick DeepSeek-VL2 for question answering over a single document or chart, where it has published DocVQA and ChartQA scores. Watch DeepSeek-OCR for looping on table headers and for 'fixing' unusual words, and DeepSeek-VL2 for blurry photos and multi-image requests.

DeepSeek-OCRdeepseek/deepseek-ocrDeepSeek-VL2deepseek/deepseek-vl2
TypeOpen sourceVision model
Price1 credit / page2 credits / page
Per 1,000 pages$1.00$2.00
Typical speed1.1 s / page3.0 s / page
Best forCheap bulk conversionCharts and tables
InputsPDF, imagesimages
Languages100+ languagesMainly English and Chinese
Max pages per request200500
Tables✓ Yes✓ Yes
HandwritingNoNo
Math✓ YesNo
Bounding boxes✓ Yes✓ Yes
ConfidenceNoNo
FormsNoNo
Multilingual✓ YesNo
On OCRRouterAvailableAvailable

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

  • It is MIT licensed, so the weights can be used commercially without a separate model licence.
  • It gives 64 to 400 vision tokens per page and about 2,500 tokens per second on one A100-40G, as quoted by the authors, which suits bulk PDF jobs.
  • It outputs Markdown, plain text, HTML and bounding boxes, so you get layout and not only text.

Choose DeepSeek-VL2 if

  • It answers questions about documents and charts, with a DocVQA test score of 93.3 and a ChartQA test score of 86.0 from the authors.
  • The Small model scores 834 on OCRBench, above the full model's 811, so a smaller model can be enough for text reading.
  • The Tiny model runs under 40GB of GPU memory, and only 1.0B to 4.5B parameters are active per token.

Key differences

DeepSeek-OCRDeepSeek-VL2
LicenceMITMIT for code; DeepSeek Model License for weights, commercial use allowed
Output formatsMarkdown, plain text, HTML, bounding boxesPlain text, Markdown
HardwareCUDA GPU; reference setup CUDA 11.8 and PyTorch 2.6 in bfloat16; throughput quoted on an A100-40GSmall may need 80GB GPU memory, or 40GB with incremental prefilling; Tiny runs under 40GB
Published benchmarksNot publishedDocVQA test 93.3; OCRBench 834 (Small); ChartQA test 86.0
Multi-page and multi-image usePDF throughput of about 2,500 tokens per second on one A100-40G (authors)Only a few images per session; dynamic tiling turns off with more than two images
Known failureRepeated tokens and looping over table headers; corrects nonsense words instead of copying them; a blank image took 30 seconds and returned repeated garbageStruggles with blurry images
PriceFree, self-hosted (you pay for compute)Free, self-hosted (you pay for compute)

Sources

What each one is

DeepSeek-OCR

Open-source vision model that compresses pages into few tokens.

DeepSeek-OCR details

DeepSeek-VL2

DeepSeek's mixture-of-experts vision models (December 2024) with OCR and table and chart understanding.

DeepSeek-VL2 details

Try 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": ["deepseek/deepseek-vl2"],"file_url": "https://example.com/document.pdf"}

Questions

More comparisons

DeepSeek-OCR alternativesDeepSeek-VL2 alternativesBest OCR models in 2026