Mistral on Apple M4

Can I run Mistral NeMo 12B on an Mac Studio M4 Max 128GB?

Yes. Mistral NeMo 12B at Q4_K_M uses 8.8 GB of the 96.0 GB available on a Mac Studio M4 Max 128GB, and runs at about 45.5 tokens per second. You can push the context to 128K.

Runs comfortably

8.8 GB of 96.0 GB · 9%
096 GB
Weights 7.0 GB
KV cache 1.3 GB
Runtime overhead 0.5 GB

Mistral NeMo 12B at Q4_K_M leaves 87.2 GB spare on a Mac Studio M4 Max 128GB. There is room to raise the context length or move up a quantisation level.

Generation45.5tok/s
Prompt processing583tok/s
Max context128Ktokens
KV per 1K tokens0GB

Every quantisation of Mistral NeMo 12B on a Mac Studio M4 Max 128GB

Highlighted row is the highest quality that still fits at 8K context.

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 22.8 GB 24.6 GB Runs comfortably 128K 14.9
INT8 / W8A8 8.50 11.8 GB 13.6 GB Runs comfortably 128K 28.0
Q8_0 (GGUF) 8.50 11.8 GB 13.6 GB Runs comfortably 128K 28.0
FP8 (E4M3) 8.00 11.4 GB 13.2 GB Runs comfortably 128K 29.0
Q6_K 6.56 9.4 GB 11.1 GB Runs comfortably 128K 34.9
Q5_K_M 5.67 8.1 GB 9.9 GB Runs comfortably 128K 39.9
AWQ 4-bit 4.25 7.9 GB 9.7 GB Runs comfortably 128K 40.8
GPTQ 4-bit 4.25 7.9 GB 9.7 GB Runs comfortably 128K 40.8
MXFP4 4.25 7.9 GB 9.7 GB Runs comfortably 128K 40.8
Q5_K_S 5.52 7.9 GB 9.7 GB Runs comfortably 128K 40.9
Q4_K_M 4.85 7.0 GB 8.8 GB Runs comfortably 128K 45.5
Q4_K_S 4.58 6.7 GB 8.5 GB Runs comfortably 128K 47.6
Q4_0 4.55 6.6 GB 8.4 GB Runs comfortably 128K 47.8
IQ4_XS 4.25 6.3 GB 8.0 GB Runs comfortably 128K 50.4
Q3_K_M 3.91 5.8 GB 7.6 GB Runs comfortably 128K 53.8
IQ3_M 3.70 5.6 GB 7.3 GB Runs comfortably 128K 56.1
IQ3_XXS 3.06 4.7 GB 6.5 GB Runs comfortably 128K 64.4
Q2_K 2.63 4.2 GB 6.0 GB Runs comfortably 128K 71.6
IQ2_XXS 2.06 3.5 GB 5.3 GB Runs comfortably 128K 84.0
IQ1_M 1.75 3.1 GB 4.9 GB Runs comfortably 128K 92.7

Also worth checking