SYLIN

DeepSeek Flash Vision

by DeepSeek · Model · Language & code models

Experimental vision version of V4 Flash, open-sourced on Hugging Face after earlier API release; 305B parameters, about 168 GB

AccessOpen weightsNeeds~168 GBDownload168 GBSize305B parametersLicenseMIT

Evidence

  1. Sep 6
    Mostly beats non-vision V4 Flash and is said to be roughly on par with Claude Opus 4.8
  2. Aug 23
    Experimental vision model released on the DeepSeek API as deepseek-v4-flash-vision-exp, accepting mixed text and image input
  3. Aug 23
    Reported a large gain over V4-Flash on multimodal agent benchmarks, close to Opus-4.8; text ability said to match V4-Flash
  4. Aug 21
    Images are billed at V4-Flash token rates, capped at 384 tokens each; Files API is free and supports reuse by file_id
  5. Aug 21
    Reported 59.3% on DeepSWE, within error of GPT 5.6 Sol Medium (61%) at likely far lower cost, per a commenter
Show all 7
  1. Aug 21
    Misread an analog clock (5:10:45) in a commenter's test where Qwen3.8 27B came nearly right
  2. Aug 21
    Read a three-series business chart correctly, pulled exact callout numbers and reasoned about leading indicators and regime shift

Cost to run