StarCoder2 on Apple M2 Pro · 32GB
✅ Runs (tight) — StarCoder2 15B @ Q8_0
StarCoder2 15B at Q8_0 needs ~17.2 GB (weights 15.8 GB + KV 640 MB + overhead 809 MB @ 8K ctx) of your 18.4 GB usable (of 32.0 GB unified memory) — fits, but little headroom — close other apps or trim context. Expect ~6–10 tok/s (usable).
context @ Q8_0: runs to its full 16K
| quant | needs | speed |
|---|---|---|
| ✓ Q8_0 | 17.2 GB | ~6–10 tok/s (usable) |
| ✓ Q5_K_M | 11.7 GB | ~9–16 tok/s (usable) |
| ✓ Q4_K_M | 10.1 GB | ~11–18 tok/s (usable) |
| ✓ Q3_K_M | 8.4 GB | ~13–20 tok/s (usable) |
$runlocal install starcoder2:15b
for your own README — links back here
Same model, other machines: Apple M1 · Apple M1 Pro · Apple M2 · Apple M3 Pro · Apple M4 Pro · Apple M3 Max
Also runs on a Apple M2 Pro: DeepSeek-R1 (distill) · Mistral Small 3 · Qwen3 · Llama 3.1 · Qwen2.5
Best models for a Apple M2 Pro · 32GB →
Architecturally similar
These share StarCoder2's KV-cache geometry — they size memory the same way, so here's how they fit on a Apple M2 Pro · 32GB too:
| model | shares | verdict here |
|---|---|---|
| Qwen2.5-Coder | 2kv/128hd4kv/128hd | ✅ Runs (tight) |
| Yi 1.5 | 4kv/128hd | ✅ Runs (tight) |
| Qwen3 Coder 30B-A3B | 4kv/128hd | 🐢 CPU-only (slow) |
| SmolLM3 3B | 4kv/128hd | ✅ Runs comfortably |
| Yi-Coder 9B | 4kv/128hd | ✅ Runs comfortably |
Check any combo yourself: open the checker →