It is on paper, but crazy enough, both models at NVFP4 run similar speeds for decode! The reason is that much more sophisticated speculative drafting is available for 27B. I’m hoping this will come to Flash Next, but I know MoEs pose challenges with that.