DeepSeek team releases new visual compression model DeepSeek-OCR

On October 20th, the DeepSeek-AI team launched a new research result - DeepSeek-OCR, proposing an innovative method for compressing long text context through visual modality. This involves rendering the long context into an image and then feeding it to the model, allowing the context that originally required thousands or tens of thousands of text tokens to be represented with just a few hundred visual tokens, achieving efficient compression of information. It is reported that DeepSeek-OCR consists of two parts: the core encoder DeepEncoder and the decoder DeepSeek3B-MoE-A570M. DeepEncoder is designed to maintain low computational activations under high-resolution input, while achieving high compression ratios to control the number of visual tokens within a manageable range. Experiments show that when the number of text tokens does not exceed 10 times the number of visual tokens (compression ratio less than 10x), the model's OCR (text recognition) accuracy can reach 97%; even when the compression ratio is increased to 20x, the accuracy remains at about 60%, demonstrating great potential in the compression of historical document contexts and research on large language model memory mechanisms. DeepSeek-OCR also has high practical value. In the OmniDocBench test, DeepSeek-OCR surpassed the GOT-OCR2.0 of StageUp Xingchen with 100 visual tokens (256 tokens per page), and was better than MinerU2.0 of the Shanghai AI Lab with less than 800 visual tokens (average of over 6000 tokens per page). In actual production, DeepSeek-OCR can generate over 200,000 pages of large language model/visual language model training data per day on a single A100-40G graphics card.