DeepSeek-OCR (contexts optical compression)
DeepSeek-OCR renders text as images to compress long context, with 97% OCR precision at under 10x compression and ~60% at 20x.
Treats the vision encoder as a text compressor, so a few hundred vision tokens carry a page of text. Raises the idea of storing old context as images to extend effective context cheaply. Also an efficient document-OCR system generating 200K+ pages/day on one A100.
- Date
- Tuesday, 21 October 2025
- Lab
- DeepSeek
- Kind
- paper
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| OCR precision at <10x text:vision compression | 97% ~60% at 20x | authors |
Authors Haoran Wei, Yaofeng Sun, Yukun Li. Context-compression use is a proposal; not shown to work at frontier scale.
Sources
This record was checked against its sources on 6 October 2026. How we check