A user running a quantized Qwen3.6-35B-A3B model on a laptop with an 8GB RTX 3070 tested llama.cpp’s —n-cpu-moe 28 setting. After moving some expert weights to system memory, the user reported generation speed rising from about 4.0 to 32.8 tokens per second for the same model. The comparison used a 512-token input and 128-token output over three runs, under limited conditions. The setup requires a model file of about 16.5GiB and additional RAM, so the entire 35B model was not held in 8GB of VRAM. Results may vary with the device, memory bandwidth and number of active experts.