can it run
Comfortably. 384GB of memory against a 21GB working set leaves headroom for context, the OS, and anything else you're running.
weights 16.5GB + KV cache 2GB at 4k context vs ~96GB usable → full. 18.5 GB needed of 96 GB usable — headroom for context
computed roofline: 1792 GB/s × 0.46–0.76 efficiency window / 16.5GB (Q4_K_M) — bands, never points.
| model file | 16.5GB (GGUF Q4_K_M) |
| minimum memory | 21GB working set |
| this machine | 384GB VRAM (4× 96GB) |
| params | 27B |
| license | apache-2.0 |
| price | $91000 |
easy run: ollama pull qwen3.8:27b
runs with: ollama · llama.cpp · koboldcpp · lm studio
→ ASUS ExpertCenter Pro ET900N G3 (GB300) — this model on the biggest machine