DeepSeek-OCR 2 (Visual Causal Flow)
3B document OCR model whose DeepEncoder V2 reorders visual tokens by semantics instead of raster scan.
Replaces fixed left-to-right, top-to-bottom visual token order with a learned causal reordering ('Visual Causal Flow'), testing whether 2D understanding can come from two cascaded 1D causal passes. Apache 2.0.
- Date
- Tuesday, 27 January 2026
- Lab
- DeepSeek
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Real5-OmniDocBench overall | 73.01 Community-submitted entry on PaddlePaddle's Real5-OmniDocBench (photographed pages) leaderboard, listed in the HF card eval results (unverified); not a DeepSeek-reported OmniDocBench figure | third-party |
arXiv submission 2026-01-28; GitHub and HF repos created 2026-01-27.
Sources
This record was checked and corrected against its sources on 6 October 2026. How we check