run it locally
Run Qwen2.5 72B on Apple M3 Max 128GB
72.7B parameters on a 128 GB (unified memory) card. It fits at Q6_K — here's the VRAM math at every quantization, the speed to expect, and the quality you trade.
| quantization | vram needed | fits 128gb? | tokens/sec | quality |
|---|---|---|---|---|
| FP16 (full) | 174.5 GB | ✗ no | — | Reference quality — no quantization loss. |
| Q8_0 | 92.5 GB | ✗ no | — | Near-lossless; rarely worth the extra space over Q6. |
| Q6_Krecommended | 71.5 GB | ✓ yes | ~6 | Virtually indistinguishable from full precision. |
| Q5_K_M | 60.2 GB | ✓ yes | ~7 | Minor loss; an excellent quality-vs-size balance. |
| Q4_K_M | 48.9 GB | ✓ yes | ~8 | Small but measurable loss; the popular default. |
| Q3_K_M | 37.5 GB | ✓ yes | ~11 | Noticeable degradation; only when you're tight on VRAM. |
Weights = params × bytes/weight, +20% for KV cache & runtime; usable VRAM is 72% (Apple unified memory) of nameplate. Tokens/sec is a bandwidth ceiling (400 GB/s) — real throughput is lower with long context. Try other combinations →
Running Qwen2.5 72B on a Apple M3 Max 128GB
At Q6_K, Qwen2.5 72B's weights take about 72 GB, inside the Apple M3 Max 128GB's 92 GB of usable memory. Virtually indistinguishable from full precision. Generation speed is bound by memory bandwidth — the GPU reads the whole model once per token — so expect on the order of 6 tokens/sec before context overhead.
Quantization is the lever
Each step down in precision shrinks the model: FP16 needs about 174 GB, Q4_K_M about 49 GB — a 72% reduction for a small, usually acceptable quality cost. Q5_K_M and Q6_K are near-lossless if you have the headroom; drop to Q3 only when you're genuinely out of VRAM. The quantization guide covers the tradeoffs in detail.
Frequently asked questions
Can a Apple M3 Max 128GB run Qwen2.5 72B?
Yes — Qwen2.5 72B fits on a Apple M3 Max 128GB at Q6_K (72 GB of the 92 GB usable), with roughly 6 tokens/sec. Higher-precision quants need more VRAM; the table above shows each option.
How much VRAM does Qwen2.5 72B need?
Qwen2.5 72B is a 72.7B-parameter model. At FP16 that's about 174 GB; at Q4_K_M (the popular default) about 49 GB, including ~20% for the KV cache and runtime. Quantization is the main lever — see the per-quant table above.
Other combinations
Qwen2.5 72B on other GPUs: RTX 3090, RTX 4090, NVIDIA L4, RTX A6000, NVIDIA L40S, A100 40GB
Other models on the Apple M3 Max 128GB: Command R 35B, Mixtral 8x7B, Llama 3.3 70B, Mixtral 8x22B, Llama 3.1 405B
Related
- VRAM calculator — any model, quant and GPU.
- Running models locally — the hardware reality.
- All run-locally combinations →