On October 20, 2025, the DeepSeek-AI team unveiled their research paper, "DeepSeek-OCR: Contextual Optical Compression," introducing an innovative approach to compress lengthy textual contexts by leveraging visual modalities. Alongside this, they open-sourced the DeepSeek-OCR model, which boasts 3 billion parameters. This model comprises a core encoder, DeepEncoder, and a decoder, DeepSeek3B-MoE-A570M. DeepEncoder is tailored for high-resolution inputs, ensuring minimal computational activation while achieving remarkable compression ratios.
Experimental results indicate that when the quantity of text tokens remains within 10 times that of visual tokens, OCR accuracy can soar to 97%. Even when the compression ratio is escalated to 20 times, the accuracy hovers around 60%. In the OmniDocBench evaluation, DeepSeek-OCR outperformed GOT-OCR2.0 using a mere 100 visual tokens and outstripped MinerU2.0 with fewer than 800 visual tokens.
In real-world production scenarios, this model can generate over 200,000 pages of training data daily for large language models/visual language models on a single A100-40G GPU.