can it run
Comfortably. 384GB of memory against a 205GB working set leaves headroom for context, the OS, and anything else you're running.
weights 164GB + KV cache 2GB at 4k context vs ~96GB usable → no. 166.0 GB vs 96 GB usable — does not fit
computed roofline: 1792 GB/s × 0.46–0.76 efficiency window / 9.2GB (Q4_K_M) — bands, never points.
MoE: file is 164GB (that decides fit) but only ~18B of experts activate per token (~9.2GB streamed) — band below is the active-expert estimate, optimistic; community MoE rows run several× the naive bandwidth math
| model file | 164GB (GGUF Q4_K_M) |
| minimum memory | 205GB working set |
| this machine | 384GB VRAM (4× 96GB) |
| params | 320B-A18B MoE |
| license | MIT |
| price | $91000 |
easy run: no stock runtime yet
→ ASUS ExpertCenter Pro ET900N G3 (GB300) — this model on the biggest machine