DeepSeek-V4-Flash-Vision-Exp FP8 E4M3 on 4× CMP 170hx (GA100) using vLLM PP4 + DSpark k=6
Text-only: 119–162 tok/s single stream - image-input: C1: 45.3 tok/s, C2: 78.2 tok/s aggregate - Warm TTFT: 0.386
This conspiracy theory is way too plausible
Moon dev: "didn't the hedge fund build DeepSeek, and think about WHY they would do that" "what if they built it to be bad on purpose, and the reason is genuinely sinister"
This ranking just exposed something nobody expected
Moon dev: "number two is moonshot kimmy k3 and honestly that placement alone made me question everything" "deepseek is getting absolutely crushed and here's the part that makes zero sense"
Western AI giants want you to believe that top-tier intelligence requires paying massive monthly subscriptions and burning billions in compute.
But Chinese open-weights models like DeepSeek, Kimi, and GLM are already matching frontier performance for a fraction of the cost. Why
You have reached the end of the archive
All of deepseek