DeepSeek-V4 First Visual Model Open Source

Dynamic Beating AI Newsletter: DeepSeek has open-sourced DeepSeek-V4-Flash-Vision-Exp.

This is the first experimental multimodal model in the DeepSeek-V4 series, which incorporates a visual module based on the V4-Flash architecture and has gained image understanding capabilities through continued training.

The model can view screenshots, read charts, and complete Agent tasks with tools. In pure text Agent tasks, it still maintains performance on par with V4-Flash.

The model has a total of 305 billion parameters and is released under the MIT license.