Alexey Fateev
@superalesha
โกI benchmark local LLMs on 4x RTX 3090s.
exact configs, tok/s, VRAM, and what broke.
โค๏ธ - 2xDGX Spark
๐96GB VRAM | Local AI
๊ฐ์
January 2026
325 ํ๋ก์ ์ค 3.7K ํฌ
I spent 67 hours of model time to find out how much dumber 4 bit really makes Qwen3.8-27B.
FP8 vs NVFP4 vs AWQ INT4 vs GGUF Q4_K_M vs NInfer on my 4x RTX 3090. 4,800 tasks, 10,120 requests, 14.5M reasoning tokens, no token caps anywhere.
The results surprised me. Big thread, lets go ๐งต
๋ ๋ณด๊ธฐ
