Mistral on Apple M3

Can I run Mixtral 8x7B on an Mac Studio M3 Ultra 256GB?

Yes. Mixtral 8x7B at Q4_K_M uses 27.9 GB of the 192.0 GB available on a Mac Studio M3 Ultra 256GB, and runs at about 50.5 tokens per second. You can push the context to 32K.

Runs comfortably

27.9 GB of 192.0 GB · 15%
0192 GB
Weights 26.4 GB
KV cache 1.0 GB
Runtime overhead 0.5 GB

Mixtral 8x7B at Q4_K_M leaves 164.1 GB spare on a Mac Studio M3 Ultra 256GB. There is room to raise the context length or move up a quantisation level.

Generation50.5tok/s
Prompt processing928tok/s
Max context32Ktokens
KV per 1K tokens0GB

Every quantisation of Mixtral 8x7B on a Mac Studio M3 Ultra 256GB

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

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 87.0 GB 88.5 GB Runs comfortably 32K 16.4
INT8 / W8A8 8.50 46.2 GB 47.7 GB Runs comfortably 32K 30.1
Q8_0 (GGUF) 8.50 46.2 GB 47.7 GB Runs comfortably 32K 30.1
FP8 (E4M3) 8.00 43.5 GB 45.0 GB Runs comfortably 32K 31.8
Q6_K 6.56 35.7 GB 37.2 GB Runs comfortably 32K 38.3
Q5_K_M 5.67 30.8 GB 32.3 GB Runs comfortably 32K 43.8
Q5_K_S 5.52 30.0 GB 31.5 GB Runs comfortably 32K 44.9
Q4_K_M 4.85 26.4 GB 27.9 GB Runs comfortably 32K 50.5
Q4_K_S 4.58 24.9 GB 26.4 GB Runs comfortably 32K 53.1
Q4_0 4.55 24.8 GB 26.3 GB Runs comfortably 32K 53.4
AWQ 4-bit 4.25 23.5 GB 25.0 GB Runs comfortably 32K 56.1
GPTQ 4-bit 4.25 23.5 GB 25.0 GB Runs comfortably 32K 56.1
MXFP4 4.25 23.5 GB 25.0 GB Runs comfortably 32K 56.1
IQ4_XS 4.25 23.1 GB 24.7 GB Runs comfortably 32K 56.8
Q3_K_M 3.91 21.3 GB 22.8 GB Runs comfortably 32K 61.1
IQ3_M 3.70 20.2 GB 21.7 GB Runs comfortably 32K 64.1
IQ3_XXS 3.06 16.7 GB 18.2 GB Runs comfortably 32K 75.5
Q2_K 2.63 14.4 GB 15.9 GB Runs comfortably 32K 85.7
IQ2_XXS 2.06 11.3 GB 12.8 GB Runs comfortably 32K 104
IQ1_M 1.75 9.6 GB 11.1 GB Runs comfortably 32K 118

Also worth checking