Yes — with 13.6 GB to spare
Qwen2.5-Coder 14B at Q4_K_M fits your M4 · 32 GB entirely in unified memory at 8K context, at an estimated 8.1 tokens per second. Past 80K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
Fully in unified memory
8K context
Q4_K_M · 8.3 GB
Apache 2.0
Released Nov 2024
Noticeably better at whole-file edits than the 7B, still comfortable on 12 GB.
The VRAM budget
weights 8.3 GB
Weights 8.3 GB
KV cache @ 8K 1.50 GB
Runtime overhead 0.6 GB
Free 13.6 GB of 24.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 27.6 GB | 29.7 GB | — | ~2.4 | Reference | 5.7 GB over |
| Q8_0 | 14.6 GB | 16.7 GB | 46K | 4.6 | −0.1% ppl | Long context |
| Q6_K | 11.3 GB | 13.4 GB | 64K | 5.9 | −0.4% ppl | Long context |
| Q5_K_M | 9.8 GB | 11.9 GB | 72K | 6.9 | −0.8% ppl | Long context |
| Q4_K_M | 8.3 GB | 10.4 GB | 80K | 8.1 | −1.9% ppl | Recommended |
| Q3_K_M | 6.7 GB | 8.8 GB | 88K | 10 | −5.4% ppl | Long context |
| Q2_K | 5.8 GB | 7.9 GB | 94K | 12 | −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/Qwen2.5-Coder-14B-Instruct-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 24.0 GB of 32 GB. Raising it with iogpu.wired_limit_mb is possible, and risky.
03There is room to go to 80K context on this card.