Yes — with 61.6 GB to spare
Phi-4 14B at Q4_K_M fits your M2 Max · 96 GB entirely in unified memory at 8K context, at an estimated 27 tokens per second. There is room for its full 16K window.
Fully in unified memory
8K context
Q4_K_M · 8.3 GB
MIT
Released Dec 2024
Trained heavily on synthetic reasoning data. Short 16K window is its main limitation.
The VRAM budget
weights 8.3 GB
Weights 8.3 GB
KV cache @ 8K 1.56 GB
Runtime overhead 0.6 GB
Free 61.6 GB of 72.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 27.4 GB | 29.5 GB | 16K | 8.2 | Reference | Fits |
| Q8_0 | 14.5 GB | 16.7 GB | 16K | 15 | −0.1% ppl | Fits |
| Q6_K | 11.2 GB | 13.4 GB | 16K | 20 | −0.4% ppl | Fits |
| Q5_K_M | 9.7 GB | 11.9 GB | 16K | 23 | −0.8% ppl | Fits |
| Q4_K_M | 8.3 GB | 10.4 GB | 16K | 27 | −1.9% ppl | Recommended |
| Q3_K_M | 6.7 GB | 8.9 GB | 16K | 33 | −5.4% ppl | Fits |
| Q2_K | 5.7 GB | 7.9 GB | 16K | 39 | −15% ppl | Fits |
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-4-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 8.3 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 16K context on this card.