DeepSeek Releases V4-Flash-Vision-Exp Weights, Developers Can Download Weights for Local Deployment

PANews reported on August 31 that, according to Hugging Face, DeepSeek has open-sourced the weights of its multimodal model DeepSeek-V4-Flash-Vision-Exp. Built on the V4-Flash architecture, the model adds image understanding capabilities, can recognize screenshots and charts, and can execute Agent tasks in combination with tools. Its performance on pure text Agent benchmarks such as Terminal Bench is roughly on par with V4-Flash. Previously, this model series was only available through APIs, but now developers can download the weights for local deployment and integrate them using frameworks such as Transformers, vLLM, and SGLang.