DeepSeek enhances multimodal capabilities with new vision model update

On August 21st, DeepSeek announced a major update to its large model toolchain, officially releasing version 0.1.1-rc.1 of the DeepSeek Harness. This update introduces a new visual model experience called V4Flash Vision-Exp, which is built upon the existing V4Flash and V4Pro models. The previous version, 0.1.0-rc.8, had already enabled DeepSeek Harness to access multimodal core capabilities. This allowed the system to support native image requests, with key commands such as /goal and /plan now accepting mixed input of text and images. Users can now send work screenshots along with specific requirements, enabling AI agents to directly 'work by looking at images,' significantly simplifying the development and operation process. Although DeepSeek previously offered a good image recognition mode on the web, integrating it deeply into the development toolchain marks the completion of the last puzzle in multimodal processing. Developers can simply upgrade by specifying an npm command to experience this new visual model capability that combines speed and accuracy.