Yes — with 68.2 GB to spare
Phi-4-mini 3.8B at Q4_K_M fits your M2 Max · 96 GB entirely in unified memory at 8K context, at an estimated 104 tokens per second. There is room for its full 128K window.
Fully in unified memory
8K context
Q4_K_M · 2.2 GB
MIT
Released Feb 2025
MIT-licensed, dense, and unusually strong on instruction following for its size.
The VRAM budget
weights 2.2 GB
Weights 2.2 GB
KV cache @ 8K 1.00 GB
Runtime overhead 0.6 GB
Free 68.2 GB of 72.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 7.2 GB | 8.8 GB | 128K | 31 | Reference | Long context |
| Q8_0 | 3.8 GB | 5.4 GB | 128K | 59 | −0.1% ppl | Long context |
| Q6_K | 2.9 GB | 4.5 GB | 128K | 76 | −0.4% ppl | Long context |
| Q5_K_M | 2.5 GB | 4.1 GB | 128K | 88 | −0.8% ppl | Long context |
| Q4_K_M | 2.2 GB | 3.8 GB | 128K | 104 | −1.9% ppl | Recommended |
| Q3_K_M | 1.7 GB | 3.3 GB | 128K | 128 | −5.4% ppl | Long context |
| Q2_K | 1.5 GB | 3.1 GB | 128K | 149 | −15% ppl | Long context |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it.
How to run it
$ pip install mlx-lm $ mlx_lm.generate --model mlx-community/Phi-4-mini-instruct-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 2.2 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02macOS caps what the GPU may wire down at about 72.0 GB of 96 GB. Raising it with iogpu.wired_limit_mb is possible, and risky.
03There is room to go to the model's full 128K context on this card.