can it run
Only with compromises. 24GB is below the 28GB working set — expect a lower quant, offloading, or heavy swap. Possible, not pleasant.
weights 22GB + KV cache 1GB at 4k context vs ~24GB usable → tight. 23.0 of 24 GB usable — keep context modest
computed roofline: 672 GB/s × 0.48–0.72 efficiency window / 22.0GB (Q4_K_M) — bands, never points.
| model file | 22GB (GGUF Q4_K_M) |
| minimum memory | 28GB working set |
| this machine | 24GB GDDR6, 672 GB/s |
| params | 32B |
| license | apache-2.0 |
| price | $900 |
easy run: ollama pull qwen2.5vl:32b
runs with: ollama · llama.cpp · koboldcpp · lm studio
→ Mac mini M5 Pro (64GB) — cheapest machine that runs it ($1669)
→ ASUS ExpertCenter Pro ET900N G3 (GB300) — this model on the biggest machine
→ every model the Titan RTX (24GB, used) can run — full list
→ rtx-3090 — the next step up from this machine's GPU