I always forget the VRAM requirements on these MOE things

fwiw because of the relatively few activated params offloading to system RAM is quite feasible, you can see the endless amount of people doing this on r/localllama with qwen3.6 35a3b