Qwen 3.8-27B NVFP4 actually beats Q5_K_M and touches official BF16 levels, but only if you change 2 sampling params. Also almost 3x faster.

So I've been going down this rabbit hole for the past week comparing quantized Qwen3.8-27B variants on my 5090, and I stumbled onto something that genuinely surprised me. Figured I'd share since I couldn't find anyone else talking about this. The setup: I was comparing two ways to run Qwen3.8-27B lo