llama-cpp

Alibaba's Qwen runs on two RTX 3060s — but the default config is 7× slower

Reddit user hit 400 t/s on Alibaba's Qwen3 125B with two RTX 3060s. Started at 36 t/s — an 11× gap, all from one default-config trap.