Qwen3.6-35B-A3B hits 79 t/s, 128K ctx

This reads less like a model breakthrough and more like a reminder that MoE placement policy can make or break local inference. On 16GB GPUs, the difference between keeping experts pinned to CPU versus splitting them intelligently is the difference between wasting VRAM and getting a genuinely usable speedup.

• –`--cpu-moe` leaves most of the GPU underutilized; `--n-cpu-moe 20` pushes the later experts onto VRAM and roughly jumps generation from 51.2 t/s to 78.7 t/s.
• –`-np 1` matters for single-user serving because it avoids wasting memory on extra recurrent-state slots, which helps make the 128K context claim believable in a constrained setup.
• –The result is hardware-specific, but the pattern is useful: for sparse MoE models, inference tuning can matter as much as quantization choice.
• –The 9800X3D’s large L3 cache probably helps the CPU side, so this is not a universal 16GB-GPU recipe, just a strong data point for consumer desktop rigs.