DeepSeek Ships V4 Flash Vision Exp Multimodal Model

DeepSeek just put a new experimental model on its API platform. It is called DeepSeek-V4-Flash-Vision-Exp, and it adds vision to the V4 Flash family.

Early benchmarks suggest it beats base Flash on visual tasks while matching it on text and agent reasoning — and the API identifier is ready to use today.

What Happened

DeepSeek rolled out DeepSeek-V4-Flash-Vision-Exp on the DeepSeek API Platform. It scores 64.3 on Chartography and 35.0 on ZeroBench, outperforming base Flash on visual tasks.

On text and agent reasoning it matches V4 Flash, with 83.9 on Terminal Bench 2.1 and 59.3 on DeepSWE — close to flagship models like Opus-4.8. Today's DeepSeek Harness 0.1.1 release supports it out of the box, and developers can call it with model="deepseek-v4-flash-vision-exp".

Why It Matters

This is one model you can drop into visual agents, coding pipelines, and multimodal testbeds without switching away from the text reasoning you already use. If you work with V4 Flash today, the vision version is the same API with a new identifier. Bookmark it for your next multimodal project.