can it run
Comfortably. 32GB of memory against a 8GB working set leaves headroom for context, the OS, and anything else you're running.
weights 6.2GB + KV cache 2GB at 4k context vs ~32GB usable → full. 8.2 GB needed of 32 GB usable — headroom for context
computed roofline: 897 GB/s × 0.46–0.76 efficiency window / 6.2GB (Q4_K_M) — bands, never points.
| model file | 6.2GB (GGUF Q4_K_M) |
| minimum memory | 8GB working set |
| this machine | 32GB HBM2, 897 GB/s |
| params | large-v3 |
| license | apache-2.0 |
| price | $645 |
easy run: no stock runtime yet
→ every model the Tesla V100 32GB (used) can run — full list
→ RTX 3090 (24GB, used) — upgrade path machine