DeepSeek-OCR: Reducing Token Counts with Optical Context Compression

DeepSeek-OCR tackles the costly problem of processing long documents by using optical context compression to reduce token counts by 7-20 times compared to traditional text-based OCR. The system combines DeepEncoder, which compresses document images into visual tokens, with a Mixture-of-Experts decoder that reconstructs text from these compressed tokens. Trained on over 30 million PDF pages across 100+ languages, it maintains 97% accuracy at 10x compression while significantly outperforming competitors on benchmarks like OmniDocBench.