Running Qwen2.5 72B locally
Alibaba · 72B parameters · released 2024-09
Qwen2.5 at near-frontier scale. Comparable memory class to Llama 3.3 70B — around 43 GB at Q4.
Q4 — 4-bit quantized
≈ 41.8 GB
weights only, plus 2 GB context headroom rule
Q8 — 8-bit quantized
≈ 76.3 GB
weights only, plus 2 GB context headroom rule
Verdicts by hardware
| Hardware | Usable | Q4 (41.8 GB) | Q8 (76.3 GB) |
|---|---|---|---|
| RTX 4090 (24 GB) | 24 GB | Too big | Too big |
| RTX 3090 (24 GB) | 24 GB | Too big | Too big |
| RTX 4070 Ti SUPER (16 GB) | 16 GB | Too big | Too big |
| RTX 4060 Ti (16 GB) | 16 GB | Too big | Too big |
| RTX 3060 (12 GB) | 12 GB | Too big | Too big |
| RTX 4060 (8 GB) | 8 GB | Too big | Too big |
| MacBook Air M2 (8 GB unified) | 5.6 GB | Too big | Too big |
| MacBook Pro M4 Pro (48 GB unified) | 33.6 GB | Too big | Too big |
| Mac Studio M2 Max (64 GB unified) | 44.8 GB | Runs well | Too big |
| CPU only (32 GB system RAM) | 25.6 GB | Too big | Too big |
| CPU only (16 GB system RAM) | 12.8 GB | Too big | Too big |
Notes
The table uses a transparent estimate: 72B × bytes-per-weight (Q4 0.58, Q8 1.06), then compares against usable memory — full VRAM on discrete GPUs, ~70% on Apple unified memory, ~80% on CPU-only system RAM. A verdict needs the weights plus 2 GB of headroom for context and runtime buffers.
Longer context windows consume additional memory beyond the weights. If you run 32k+ token contexts, treat “Tight fit” as “no” and give “Runs well” setups extra headroom.