DeepSeek-OCR: Contexts Optical Compression скачать в хорошем качестве

DeepSeek-OCR: Contexts Optical Compression

The paper introduces DeepSeek-OCR, a vision-language model designed to explore the concept of contexts optical compression, which uses visual modality to efficiently compress long textual contexts for Large Language Models (LLMs). DeepSeek-OCR consists of the DeepEncoder and a DeepSeek3B-MoE decoder, with the DeepEncoder specifically engineered to achieve high compression ratios while managing a low number of vision tokens. Experiments on benchmarks like Fox and OmniDocBench demonstrate that the model can maintain high Optical Character Recognition (OCR) precision, even at compression ratios up to 10×, showcasing its potential to mitigate the computational challenges of processing lengthy text sequences. Beyond its theoretical research value in memory forgetting mechanisms and context compression, DeepSeek-OCR proves highly practical, achieving state-of-the-art OCR performance with fewer vision tokens than competitors and capable of generating substantial training data for other LLMs and VLMs.