On October 20, 2025, the DeepSeek-AI team unveiled a cutting-edge OCR model, DeepSeek-OCR, and made its core code open to the public. This model leverages an innovative technique known as 'Contextual Optical Compression' to effectively tackle the efficiency obstacles encountered in long-text processing scenarios, boasting a substantial parameter count of 3 billion.

The model is composed of two key components: a DeepEncoder encoder and a DeepSeek3B-MoE-A570M decoder. DeepEncoder incorporates SAM and CLIP technologies, enabling it to achieve low computational requirements and high compression ratios for high-resolution input data. Experimental results demonstrate that when the compression ratio remains below 10 times, the OCR accuracy can soar to an impressive 97%. Even when the compression ratio is pushed to 20 times, the model still maintains an accuracy rate of approximately 60%.

In the OmniDocBench testing environment, DeepSeek-OCR outshines its competitors. It surpasses GOT-OCR2.0, which necessitates 256 visual tokens, by achieving comparable performance with only 100 tokens. Additionally, it outperforms MinerU2.0, which requires over 6,000 tokens, using less than 800 tokens.

DeepSeek-OCR holds significant value in both research and practical applications. It supports research on historical document compression and the memory mechanisms of large models. Moreover, a single A100-40G GPU can generate over 200,000 pages of training data per day. The model is also capable of parsing complex elements such as charts and molecular formulas, making it an efficient tool for digitization efforts in fields like finance and scientific research. The relevant code and model weights are now freely accessible on GitHub and Hugging Face.