gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090-LMHead4

Mark superseded by the final main checkpoint

NVFP4 lm_head variant: -8.8% size, +8.4% decode, no accuracy cost

initial commit