can it run
Comfortably. 32GB 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 ~32GB usable → full. 5.8 GB needed of 32 GB usable — headroom for context
computed roofline: 1792 GB/s × 0.55–0.80 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 | 32GB GDDR7 |
| params | 8B |
| license | apache-2.0 |
| price | $5000 |
easy run: ollama pull qwen3:8b
runs with: ollama · llama.cpp · koboldcpp · lm studio
→ every model the RTX 5090 (32GB) can run — full list
→ RTX PRO 6000 Blackwell (96GB) — upgrade path machine
B&H Photo ↗ · Amazon ↗ · Amazon ↗