Qwen2.5 on Apple M4 Pro · 48GB
✅ Runs (tight) — Qwen2.5 72B @ Q2_K
Qwen2.5 72B at Q2_K needs ~32.3 GB (weights 28.4 GB + KV 2.5 GB + overhead 1.4 GB @ 8K ctx) of your 33.0 GB usable (of 48.0 GB unified memory) — fits, but little headroom — close other apps or trim context. Expect ~4–7 tok/s (slow).
context @ Q2_K: runs to 32K · won't fit past that on this rig
| quant | needs | speed |
|---|---|---|
| ✗ Q5_K_M | 53.1 GB | — |
| ✗ Q4_K_M | 45.6 GB | — |
| ~ Q3_K_M | 37.2 GB | ~0.5–0.8 tok/s (slow) |
| ✓ Q2_K | 32.3 GB | ~4–7 tok/s (slow) |
$runlocal install qwen2.5:72b
for your own README — links back here
Same model, other machines: Apple M1 · Apple M1 Pro · Apple M2 · Apple M3 Pro · Apple M2 Pro · Apple M3 Max
Also runs on a Apple M4 Pro: Qwen3 · DeepSeek-R1 (distill) · Mixtral 8x7B · Gemma 3 · Mistral Small 3
Best models for a Apple M4 Pro · 48GB →
Architecturally similar
These share Qwen2.5's KV-cache geometry — they size memory the same way, so here's how they fit on a Apple M4 Pro · 48GB too:
| model | shares | verdict here |
|---|---|---|
| Llama 3.1 | 8kv/128hd | ✅ Runs (tight) |
| Qwen2.5-Coder | 2kv/128hd4kv/128hd8kv/128hd | ✅ Runs comfortably |
| Llama 3.2 | 8kv/128hd | ✅ Runs comfortably |
| Llama 3.3 | 8kv/128hd | ✅ Runs (tight) |
| Mistral 7B | 8kv/128hd | ✅ Runs comfortably |
| Mistral Nemo | 8kv/128hd | ✅ Runs comfortably |
| Codestral | 8kv/128hd | ✅ Runs comfortably |
| Llama 3.2 Vision | 8kv/128hd | ✅ Runs comfortably |
…and 11 more — see the Qwen2.5 page.
Check any combo yourself: open the checker →