Mantrah/Qwen3.8-27B-NVFP4-GDN

Model card: long-context RULER-lite vs unsloth (needle 64/64 both, VT-120K 10 vs 13), gsm8k with official sampling (95.5 vs 96.4), greedy note, 460 W power limit

Model card: correct the novelty claim — GDN-in-NVFP4 uploads existed since 08-14; what this adds is the FP8 lm_head + side-by-side measurements

Qwen3.8-27B NVFP4, linear attention included

initial commit