Qwen3.8-Flash-Next IQ3_S on one RTX PRO 6000: fastest published decode, near-total KV reuse at 80K
Qwen3.8-Flash-NextReproducible Strata 0.1.56 setup with DFlash2 speculation and parked-KV sweeps: rank 1 and 2 of 806 runs, 99.95% prompt-token reuse at every context length up to 80K.
