can it run
Comfortably. 32GB of memory against a 6GB working set leaves headroom for context, the OS, and anything else you're running.
weights 4.7GB + KV cache 2GB at 4k context vs ~32GB usable → full. 6.7 GB needed of 32 GB usable — headroom for context
computed roofline: 897 GB/s × 0.46–0.76 efficiency window / 4.7GB (Q4_K_M) — bands, never points.
| model file | 4.7GB (GGUF Q4_K_M) |
| minimum memory | 6GB working set |
| this machine | 32GB HBM2, 897 GB/s |
| params | 7B |
| license | apache-2.0 |
| price | $645 |
easy run: ollama pull llava:7b
runs with: ollama · llama.cpp · koboldcpp · lm studio
→ every model the Tesla V100 32GB (used) can run — full list
→ RTX 3090 (24GB, used) — upgrade path machine