DeepSeek OCR AI model can process 200,000 document pages a day on a single Nvidia A100 GPU

DeepSeek has launched an open-source OCR model that processes over 200,000 document pages daily on a single Nvidia A100 GPU by converting text into compressed visual tokens. The system achieves 97% recognition accuracy while using nine times fewer computing resources than traditional methods by converting multiple text tokens into single visual tokens. DeepSeek-OCR outperforms competing solutions and can handle complex multilingual documents with scientific formulas and diagrams, potentially transforming how AI models learn from text-heavy content.