DeepSeek-OCR (Contexts Optical Compression)
Renders text as images to compress context, reaching 97% decoding accuracy at under 10x compression and about 60% at 20x with a 3B-MoE decoder.
DeepEncoder compresses a page to as few as 64 vision tokens; a DeepSeek3B-MoE-A570M decoder reads them back. Proposes images as a compressed memory medium for LLMs; generates 200K+ pages/day of training data per A100-40G.
- Date
- Monday, 20 October 2025
- Lab
- DeepSeek
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Decoding precision at <10x compression | 97% About 60% at 20x; 200K+ pages/day on one A100-40G | company |
Release date per GitHub (2025-10-20); arXiv submission 2025-10-21; HF repo created 2025-10-17.
Sources
This record was checked against its sources on 6 October 2026. How we check