Breaking

DeepSeek Adds Image Understanding 'Vision' to Its V4 Lineup

June 18, 2026 at 02:56 EDT

Foundation Models
Open Source

On June 18, 2026, DeepSeek added an image-understanding feature called "Vision" to its "DeepSeek-V4" model series, making it available on the web (chat.deepseek.com) and in its app. The move integrates multimodal capability into a V4 that had been text-focused, and was announced by Xiaokang Chen, a researcher on the Multimodal team.

June 18, 2026 · DeepSeek

DeepSeek-V4 goes multimodal with 'Vision' — cheap image input arrives

Image-input support is now live on web and app, built on the lightweight V4-Flash base. It keeps a 1M-token context as standard while driving the cost of seeing images far below rivals.

~90

KV cache entries per image

10×

cheaper vs Claude Vision

token context, standard

KV cache entries per image

Lower is cheaper — DeepSeek Vision holds roughly a tenth of Claude's.

~90

DeepSeek
Vision

~870

Claude
Vision

Inference FLOPs at 1M context

27%

of DeepSeek-V3.2

KV cache at 1M context

10%

of DeepSeek-V3.2

The V4 series

DeepSeek-V4-Pro

1.6T total / 49B active

Agentic coding & long-context reasoning

DeepSeek-V4-Flash

284B total / 13B active

Lightweight base for Vision support

WELCOMED BY DEVELOPERS

Replaces sub-agents & external Vision tools
Spots even small details inside images
Price edge + 1M context suits agentic use

CAVEATS

V4-Pro may still appear text-only
Must switch to a Vision-capable model
A unified-Vision successor still awaited

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.