can it run
Comfortably. 256GB of memory against a 82GB working set leaves headroom for context, the OS, and anything else you're running.
weights 65GB + KV cache 0.28GB at 4k context vs ~32GB usable → no. 65.3 GB vs 32 GB usable — does not fit
computed roofline: 1792 GB/s × 0.55–0.80 efficiency window / 9.8GB (Q4_K_M) — bands, never points.
MoE: file is 65GB (that decides fit) but only ~18B of experts activate per token (~9.8GB streamed) — band below is the active-expert estimate, optimistic; community MoE rows run several× the naive bandwidth math
| model file | 65GB (GGUF Q4_K_M) |
| minimum memory | 82GB working set |
| this machine | 256GB VRAM (8× 32GB) |
| params | 120B MoE |
| license | apache-2.0 |
| price | $100000 |
easy run: ollama pull gpt-oss:120b
runs with: ollama · llama.cpp · koboldcpp · lm studio
→ ASUS ExpertCenter Pro ET900N G3 (GB300) — this model on the biggest machine
→ rtx-pro-6000 — the next step up from this machine's GPU