can it run
Comfortably. 384GB of memory against a 7GB working set leaves headroom for context, the OS, and anything else you're running.
weights 5.2GB + KV cache 0.56GB at 4k context vs ~96GB usable → full. 5.8 GB needed of 96 GB usable — headroom for context
computed roofline: 1792 GB/s × 0.46–0.76 efficiency window / 5.2GB (Q4_K_M) — bands, never points.
| model file | 5.2GB (GGUF Q4_K_M) |
| minimum memory | 7GB working set |
| this machine | 384GB VRAM (4× 96GB) |
| params | 8B |
| license | apache-2.0 |
| price | $91000 |
easy run: ollama pull qwen3:8b
runs with: ollama · llama.cpp · koboldcpp · lm studio
no alternatives — every machine in the catalog can run this one.