Reading pictures

Vision on DeepSeek, today, means you send a picture and get text back. The experimental model name is deepseek-v4-flash-vision-exp. I used it on an ops chart: it named the series, spotted a Tuesday dip, and warned that the axis was truncated. That is useful. It is not a drawing tool.

Limits I copied from the docs: user messages only, format sniffing from bytes, size caps, and a token conversion after resize. Other V4 models reject images.

Model page: DeepSeek-V4-Flash-Vision-Exp. Text-only work stays on Flash.