DeepSeek launches image recognition mode and supports visual CoT inference based on the primitive framework of recall According to Beating monitoring, DeepSeek's web and app platforms have officially launched Vision Mode, which is available alongside Quick Mode and Expert Mode above the dialogue input box. The newly launched visual comprehension ability is not simply text recognition (OCR), but focuses on deep scene analysis, spatial logic reasoning, and directly converting UI interface screenshots into HTML structured code. For high difficulty geometric deduction or complex chart analysis, the system will automatically activate the deep thinking model and provide a complete reasoning chain. The underlying pattern of image recognition is based on the research framework "Thinking with Visual Primitives" published by the DeepSeek team. A paper jointly published by multimodal researcher Xiaokang Chen, Peking University, and Tsinghua University points out that existing visual language models have a "reference gap" in fine localization and spatial reasoning, which makes it difficult to describe complex visual coordinates using fuzzy natural language. To this end, the research team elevated the coordinate points and bounding boxes to the minimum thinking unit, and directly inserted spatial primitives into the visual reasoning chain (CoT) of the model, achieving synchronous spatial pointing during the thinking process. Academic papers and open-source projects that serve as the foundation of visual abilities were briefly released on April 30th, but were immediately withdrawn by DeepSeek officials without warning on May 1st, sparking many speculations in the industry about excessive leakage of technical details and subsequent optimization of models. The officially launched image recognition mode only supports image input and currently does not support multimodal formats such as video and audio, and the model currently does not have image generation capabilities. [Original link]