Yes — with 15.3 GB to spare
Mistral NeMo 12B at Q4_K_M fits your M2 Pro · 32 GB entirely in unified memory at 8K context, at an estimated 16 tokens per second. Past 105K 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 · 6.9 GB
Apache 2.0
Released Jul 2024
Multilingual 12B with a 128K window, built with NVIDIA. A roleplay and fiction favourite that refuses to die.
The VRAM budget
weights 6.9 GB
Weights 6.9 GB
KV cache @ 8K 1.25 GB
Runtime overhead 0.6 GB
Free 15.3 GB of 24.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 22.7 GB | 24.6 GB | 4K | ~4.9 | Reference | 0.6 GB over |
| Q8_0 | 12.1 GB | 13.9 GB | 72K | 9.3 | −0.1% ppl | Long context |
| Q6_K | 9.3 GB | 11.2 GB | 90K | 12 | −0.4% ppl | Long context |
| Q5_K_M | 8.1 GB | 9.9 GB | 98K | 14 | −0.8% ppl | Long context |
| Q4_K_M | 6.9 GB | 8.7 GB | 105K | 16 | −1.9% ppl | Recommended |
| Q3_K_M | 5.6 GB | 7.4 GB | 114K | 20 | −5.4% ppl | Long context |
| Q2_K | 4.8 GB | 6.6 GB | 119K | 23 | −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/Mistral-Nemo-Instruct-2407-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 6.9 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 105K context on this card.