can it run
Comfortably. 24GB of memory against a 4GB working set leaves headroom for context, the OS, and anything else you're running.
weights 2.5GB + KV cache 0.5GB at 4k context vs ~24GB usable → full. 3.0 GB needed of 24 GB usable — headroom for context
computed roofline: 672 GB/s × 0.48–0.72 efficiency window / 2.5GB (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 | 2.5GB (GGUF Q4_K_M) |
| minimum memory | 4GB working set |
| this machine | 24GB GDDR6, 672 GB/s |
| params | mini 3.8B |
| license | mit |
| price | $900 |
easy run: ollama pull phi4-mini
runs with: ollama · llama.cpp · koboldcpp · lm studio
→ every model the Titan RTX (24GB, used) can run — full list
→ rtx-3090 — the next step up from this machine's GPU