Unsloth claims 2.5x faster NVFP4 quants for Qwen3.6

Unsloth's NVFP4 quantizations achieve 2.5x speedup on Qwen3.6 27B and 1.56-1.79x on 35B-A3B compared to NVIDIA's official quants. The method uses W4A4 4-bit tensor cores and FP8 KV cache calibration with no accuracy loss.