can it run
Comfortably. 16GB of memory against a 7GB working set leaves headroom for context, the OS, and anything else you're running.
weights 5.6GB + KV cache 0.22GB at 4k context vs ~16GB usable → full. 5.8 GB needed of 16 GB usable — headroom for context
computed roofline: 900 GB/s × 0.46–0.76 efficiency window / 5.6GB (Q4_K_M) — bands, never points.
| model file | 5.6GB (GGUF Q4_K_M) |
| minimum memory | 7GB working set |
| this machine | 16GB HBM2, 900 GB/s |
| params | 7B |
| license | apache-2.0 |
| price | $340 |
easy run: ollama pull qwen2.5vl:7b
runs with: ollama · llama.cpp · koboldcpp · lm studio
→ every model the Tesla V100 (16GB, used) can run — full list
→ RTX 3090 (24GB, used) — upgrade path machine