can it run
Only with compromises. 16GB is below the 21GB working set — expect a lower quant, offloading, or heavy swap. Possible, not pleasant.
weights 16.5GB + KV cache 2GB at 4k context vs ~16GB usable → offload-partial. 18.5 GB vs 16 GB usable — partial CPU offload, expect large speed loss
computed roofline: 900 GB/s × 0.46–0.76 efficiency window / 16.5GB (Q4_K_M) — bands, never points.
| model file | 16.5GB (GGUF Q4_K_M) |
| minimum memory | 21GB working set |
| this machine | 16GB HBM2, 900 GB/s |
| params | 27B |
| license | apache-2.0 |
| price | $340 |
easy run: ollama pull qwen3.8:27b
runs with: ollama · llama.cpp · koboldcpp · lm studio
→ Mac mini M4 (32GB) — cheapest machine that runs it ($999)
→ ASUS ExpertCenter Pro ET900N G3 (GB300) — this model on the biggest machine
→ every model the Tesla V100 (16GB, used) can run — full list
→ RTX 3090 (24GB, used) — upgrade path machine