DeepSeek Open Sources Again: Releases 3B MoE OCR Model DeepSeek-OCR

DeepSeek introduces its new visual text compression model, DeepSeek-OCR. With only 3 bytes of parameters, a hybrid expert architecture, and a 20x reduction in visual tokens, it achieves a 20x compression ratio and can process 33 million pages of data per day across 20 nodes. In Fox benchmark tests, it achieves over 85% accuracy across various text length ranges. It supports multiple resolution configurations, multilingual processing, complex chart parsing, and other multimodal capabilities, achieving 10x compression efficiency in multi-turn dialogues.