Unsloth claims 2.5x faster NVFP4 quants for Qwen3.6

Unsloth reports 2.5x speedup for Qwen3.6 27B and 1.56-1.79x for 35B-A3B with NVFP4 quantizations vs NVIDIA's. Their approach uses W4A4 tensor cores vs NVIDIA's W4A16, with no accuracy loss.