Qwen3.8-Flash-Next, on 5090+64gb, with Llama.cpp - Seems to not use ram?

I was actually pretty happy with my Qwen3.8-27b setup, and I'd been tinkering with Ninfer to have a version that was "fast but maybe a bit stupid" and the speed was nice to have as a backup. But I was curious how the Flash-Next version might work, after I learned it didn't need to all fit in VRAM to work. I picked up the Atomic quant (let me know if there is a better one I should use, this one seemed good from what I could find). I used the build setup below. It can still be tweaked some more, as I still am only using about 27gb of my vram. ./build/bin/llama-server \ --model "/mnt/SPCC-2TB/Projects/AI-APPS/LLM-Models/Qwen3.8-Flash-Next-Atomic/Qwen3.8-Flash-Next-AD-4.27bpw-Q4_K_M-M64 -00001-of-00033.gguf" \ --no-mmproj \ --load-mode mmap \ --lazy-mode on \ --fit off \ --gpu-layers all \ --n-cpu-moe 32 \ --ctx-size 64768 \ --flash-attn on \ --jinja \ --parallel 1 The odd thing I noticed though - I know that some parts of this are meant to run from the SSD for the sake of saving vram space etc. Fine. But I kinda expected that some of it at least would get buffered into system ram, as running from ram would be a whole lot more efficient than running from my NVME drive. But this run gets me the following results: 40tok/s decode. 50tok/s prompt processing (It's a short prompt so probably not accurate) 27gb of vram used 8gb of system ram used... So... I mean, am I just wrong and this is normal? The speed doesn't seem as bad as I expected (I thought I was going to get more like 10tok/s at best) but it seems like I might be missing a trick somewhere?