can it run
Comfortably. 24GB of memory against a 10GB working set leaves headroom for context, the OS, and anything else you're running.
weights 8GB + KV cache 3.12GB at 4k context vs ~24GB usable → full. 11.1 GB needed of 24 GB usable — headroom for context
computed roofline: 288 GB/s × 0.28–0.50 efficiency window / 8.0GB (Q4_K_M) — bands, never points.
| model file | 8.0GB (GGUF Q4_K_M) |
| minimum memory | 10GB working set |
| this machine | 24GB GDDR5, 288 GB/s |
| params | 13B |
| license | apache-2.0 |
| price | $55 |
easy run: ollama pull llava:13b
runs with: ollama · llama.cpp · koboldcpp · lm studio
→ every model the Tesla M40 (24GB, used) can run — full list
→ tesla-p40 — the next step up from this machine's GPU