Qwen3.8 flash next + exllamav3 + hermes is amazing

A user reports that the Qwen3.8 flash next model, combined with exllamav3 and the Hermes agent, delivers impressive performance. Running Qwen3.8fn 6bpw exl3 (from turboderp) on a 6x3090 setup yields 80-120 tokens per second with good quality. The user highlights the engine's efficiency for CUDA, suggesting that lower quantizations on fewer GPUs might achieve similar results.