DeepSeek Adds a Visual Reasoning Mode to Its Chatbot

DeepSeek has rolled out a Vision Mode across its web and mobile apps, according to the AI ML Big Data channel. The feature applies visual chain-of-thought reasoning to image inputs. The company describes its target tasks as geometric reasoning, chart and graph analysis, and the direct conversion of user-interface screenshots into HTML.

The screenshot-to-HTML capability is the most concrete claim for developers. A model that can take a rendered interface and produce working markup is directly useful for front-end work and for agent pipelines that operate over graphical applications. The geometry and chart-reading targets point at the same weakness most multimodal systems still show, which is reasoning over precise spatial structure rather than describing an image in general terms.

The report attributes the underlying capability to DeepSeek's reasoning model line. As with any vendor-announced feature, the framing is the company's own, and there are no published benchmark numbers, no named evaluation, and no independent reproduction yet. The useful test is not the demonstration gallery but whether the mode holds up on difficult diagrams and dense interfaces where competing multimodal models tend to fail.