DeepSeek Releases V4-Flash-Vision-Exp Weights for Local Deployment
- DeepSeek has open-sourced the weights of its multimodal model DeepSeek-V4-Flash-Vision-Exp on Hugging Face.
- The model is built on the V4-Flash architecture and adds image understanding capabilities including screenshot and chart recognition.
- The model can execute Agent tasks in combination with tools and performs comparably to V4-Flash on pure text Agent benchmarks such as Terminal Bench.
- Previously this model series was only available through APIs, but developers can now download the weights for local deployment.
- The weights can be integrated using frameworks such as Transformers, vLLM, and SGLang.
DeepSeek has open-sourced the weights of its multimodal model DeepSeek-V4-Flash-Vision-Exp, making them available for download on Hugging Face. The model builds upon the V4-Flash architecture and incorporates image understanding capabilities, enabling it to recognize screenshots and charts while executing Agent tasks in combination with tools.
Performance-wise, the model demonstrates comparable results to V4-Flash on pure text Agent benchmarks such as Terminal Bench. A significant shift in accessibility has occurred, as this model series was previously available only through API access.
Developers can now download the weights for local deployment and integrate them using popular frameworks including Transformers, vLLM, and SGLang, providing greater flexibility for implementation and customization.
DeepSeek开源V4-Flash-Vision-Exp模型权重，支持本地部署
DeepSeek已开源其多模态模型DeepSeek-V4-Flash-Vision-Exp的权重，并在Hugging Face上提供下载。该模型基于V4-Flash架构构建，新增了图像理解能力，可识别截图和图表，并能与工具结合执行Agent任务。
在性能表现上，该模型在Terminal Bench等纯文本Agent基准测试中与V4-Flash性能相当。该模型系列的可访问性发生了重大转变，此前仅通过API提供。
开发者现可下载权重进行本地部署，并使用Transformers、vLLM和SGLang等流行框架进行集成，为实现和定制提供了更大的灵活性。