原帖内容

Made A Recipe To Give Deepseek V4 FLASH Vision ! It works with the current Deepseek Model Your Running. DeepSeek V4 Flash can't see. Now it can. I bolted a 0.8B vision model onto it and put a shim in front that speaks the normal OpenAI API. Point any harness at one URL and your text-only model just accepts images. No tool wiring. No preprocessing. No harness patch. No redownloads I sent it a photo. It came back: "A boy in an orange shirt is pouring green liquid from a bowl over his own head, laughing as it splashes around him." That's a model with zero vision, describing a photograph. 1.7 seconds. Costs about 3GB of RAM next to DS4. On 2x DGX Spark you've got ~18GB spare, so it fits with room left. Zero dependencies. Pure stdlib. MIT. http://github.com/tonyd2wild/DeepSeek-v4-Flash-0731-Vision-DSpark-1M-NVFP4-KV-2x-DGX-Spark