Deepseek V4 Flash Vision Experimental is live

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

Officially available on Api. Open weights could be dropped any moment.
I expect it to be vision aware weights, not vision tower + projector.

Screenshot_20260821_141456_Reddit
Screenshot_20260821_141456_Reddit1779×1466 315 KB

DeepSeek-V4-Flash-Vision-Exp is now live on the DeepSeek API Platform!

  • This experimental multimodal model matches DeepSeek-V4-Flash on text capabilities—including agents, reasoning, and world knowledge.
  • On multimodal agent benchmarks, V4-Flash-Vision-Exp makes a major leap over V4-Flash, bringing multimodal agent performance close to Opus-4.8.
  • Try it with model=‘deepseek-v4-flash-vision-exp’. DeepSeek Harness 0.1.1 was released today with out-of-the-box support for the new model.
3

Finally! Glad to see there’s some improvements on the other benchmarks too. Deepseek v4 flash is the gift that keeps on giving

4

Indeed, that that everyone on LocalLLama reddit is screaming and moaning about 6x price hike we can finally keed nodding about our smart financial choices :D

5

It’s def going to drop soon. I almost have my vision sidecar working well. That’s how I know.

6

Anyone got a link for where it is going to land?

7

DeepSeek released all of their open weight models on Hugging Face, so deepseek-ai (DeepSeek) should be a good place to check regularly

8

Wonder if someone will make a single spark version, too. If they can make it similar in quality to those EXL quants in the other DSV4F topic, it might finally be a full replacement for 3.5 122B in every way.