can it run
Comfortably. 16GB of memory against a 4GB working set leaves headroom for context, the OS, and anything else you're running.
weights 2.6GB + KV cache 0.56GB at 4k context vs ~16GB usable → full. 3.2 GB needed of 16 GB usable — headroom for context
computed roofline: 960 GB/s × 0.50–0.80 efficiency window / 2.6GB (Q4_K_M) — bands, never points. Small models (under ~4.5B) run BELOW these bands — kernel-launch overhead keeps them from saturating bandwidth (measured against community medians, 09-06).
| model file | 2.6GB (GGUF Q4_K_M) |
| minimum memory | 4GB working set |
| this machine | 16GB GDDR7 |
| params | 4B |
| license | apache-2.0 |
| price | $1400 |
easy run: ollama pull qwen3:4b
runs with: ollama · llama.cpp · koboldcpp · lm studio
→ every model the RTX 5080 (16GB) can run — full list
→ rtx-5090 — the next step up from this machine's GPU
Amazon ↗ · Newegg ↗ · Amazon ↗ (affiliate)