can it run
Only with compromises. 192GB is below the 205GB working set — expect a lower quant, offloading, or heavy swap. Possible, not pleasant.
weights 164GB + KV cache 2GB at 4k context vs ~190GB usable → tight. 166.0 of 190 GB usable — keep context modest
computed roofline: 256 GB/s × 0.60–0.85 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 | 192GB LPDDR5X-8533 unified (160GB GPU-assignable) |
| params | 320B-A18B MoE |
| license | MIT |
| price | $None |
easy run: no stock runtime yet
→ ASUS ExpertCenter Pro ET900N G3 (GB300) — cheapest machine that runs it ($60000)
→ ASUS ExpertCenter Pro ET900N G3 (GB300) — this model on the biggest machine
HotHardware (announce) ↗ · Amazon ↗