DeepSeek V4 Flash 0731 on Strix Halo: draft model, n_max sweep, and a launch line that actually helps

I know the front page is all Qwen3.8 27B right now. Fine model. I just finished a week of numbers on DeepSeek-V4-Flash-0731 on a Strix Halo box and figured I'd dump them here before I forget what I actually measured. Hardware is a Strix Halo mini PC, Ubuntu, 128 GB unified memory. Target is Unsloth'