I've been running LLMs on my old potato i5-8500 with 32GB of RAM and *no GPU* for awhile now, running up to 12B dense models which run slow but perfectly useable. But this Gemma-4-26B-A4B simply flies on this CPU - only machine using Koboldcpp on Linux. That's right, an old used $150 desktop computer is running state of the art LLMs with something like 7 T/s. Yeah, go ahead and scoff. You can brag about your super-rig that costs more than a used car, but I'm bragging about a crappy old desktop...