Qwen3.6 larger quants are faster

Running Qwen3.6-35B-A3B-UD on an RTX 3070 8GB and DDR4 64GB setup, larger quants performed better than expected.

• IQ4_XS.gguf, at about 18GB, recorded 25~30 tok/s at 32k context.
• The larger Q4_K_XL.gguf, at about 23GB, was faster at 32 tok/s with 128k context.
• In the end, Q5_K_S offered the best balance of quality and speed, delivering about 30 tok/s.
• Even at 50k context, speed stayed above 25 tok/s.

The author concluded that for MoE models, even when VRAM looks tight, a higher quant than expected can be the better choice.