Phi-4-mini on Apple M2 · 16GB

✅ Runs comfortably — Phi-4-mini 3.8B @ Q8_0
Phi-4-mini 3.8B at Q8_0 needs ~5.3 GB (weights 3.8 GB + KV 1.0 GB + overhead 512 MB @ 8K ctx) of your 8.2 GB usable (of 16.0 GB unified memory) — plenty of headroom. Expect ~12–20 tok/s (usable).
context @ Q8_0: comfortable to 16K · runs to 64K · won't fit past that on this rig
quantneedsspeed
✓ Q8_05.3 GB~12–20 tok/s (usable)
✓ Q5_K_M4.0 GB~16–25 tok/s (fast)
✓ Q4_K_M3.7 GB~18–30 tok/s (fast)
$runlocal install phi4-mini:3.8b
runlocal verdict card
download card share on x share on reddit drop it in your model-drop post
Phi-4-mini: Runs comfortably
for your own README — links back here

Same model, other machines: Apple M1 · Apple M1 Pro · Apple M3 Pro · Apple M2 Pro · Apple M4 Pro · Apple M3 Max

Also runs on a Apple M2: DeepSeek-R1 (distill) · Qwen3 · Llama 3.1 · Qwen2.5 · Gemma 3

Best models for a Apple M2 · 16GB →

Architecturally similar

These share Phi-4-mini's KV-cache geometry — they size memory the same way, so here's how they fit on a Apple M2 · 16GB too:

modelsharesverdict here
Qwen2.5-Coder 8kv/128hd ✅ Runs (tight)
Llama 3.2 8kv/128hd ✅ Runs comfortably
Llama 3.3 8kv/128hd ❌ Won't fit
Mistral 7B 8kv/128hd ✅ Runs (tight)
Mistral Small 3 8kv/128hd 🐢 CPU-only (slow)
Mistral Nemo 8kv/128hd 🐢 CPU-only (slow)
Codestral 8kv/128hd 🐢 CPU-only (slow)
Llama 3.2 Vision 8kv/128hd ✅ Runs (tight)

…and 7 more — see the Phi-4-mini page.

Check any combo yourself: open the checker →