DeepSeek-OCR 2
DeepSeek-OCR 2 is DeepSeek's second open-source OCR model. Its new DeepEncoder V2 reads a page in an order that follows the content instead of a fixed scan.
A release of DeepSeek-OCR.
- Released
- January 2026
- Model ID
- deepseek-ai/DeepSeek-OCR-2
- Size
- About 3.4B in total: a 3B mixture-of-experts decoder with about 500M active parameters
- License
- Apache-2.0 (DeepSeek-OCR was MIT)
- Also searched as
- deepseek ocr 2, deepseek-ocr-2
What is new in DeepSeek-OCR 2
- DeepEncoder V2 reorders visual tokens by what the image contains ("Visual Causal Flow") instead of a rigid raster scan.
- The CLIP part of the encoder is replaced by a compact language-model encoder initialised from Qwen2-0.5B. The decoder is unchanged.
- 3.73 points higher on OmniDocBench v1.5 than DeepSeek-OCR with similar training data.
- Reading-order edit distance falls from 0.085 to 0.057.
- Each page uses between 256 and 1,120 visual tokens.
DeepSeek-OCR 2 benchmarks
| Benchmark | Score | Note |
|---|---|---|
| OmniDocBench v1.5, overall | 91.09% | 3.73 points above DeepSeek-OCR |
| OmniDocBench v1.5, reading-order edit distance | 0.057 | DeepSeek-OCR: 0.085. Lower is better |
Scores as published by the authors. We have not rerun them.
Use DeepSeek-OCR 2 through an API
OCRRouter runs the first DeepSeek-OCR as deepseek/deepseek-ocr. DeepSeek-OCR 2 is not available through OCRRouter yet.
1PaddleOCR-VLOpen-source document model from Baidu. Excellent on Chinese, Japanese and Korean.
paddle/paddleocr-vl6 credits / page2GLM-OCRSmall open-source OCR model from Z.ai with layout regions. MIT licensed.zai/glm-ocr6 credits / pageDeepSeek-OCR versions
- DeepSeek-OCR: all versions, price, speed and features on OCRRouter
- DeepSeek-OCR 2 (January 2026), this page
Questions
Sources
Last updated .