Cohere on Apple M3

Can I run Command R 35B on an Mac Studio M3 Ultra 256GB?

Yes. Command R 35B at Q4_K_M uses 30.7 GB of the 192.0 GB available on a Mac Studio M3 Ultra 256GB, and runs at about 18.8 tokens per second. You can push the context to 128K.

Runs comfortably

30.7 GB of 192.0 GB · 16%
0192 GB
Weights 20.1 GB
KV cache 10.0 GB
Runtime overhead 0.6 GB

Command R 35B at Q4_K_M leaves 161.3 GB spare on a Mac Studio M3 Ultra 256GB. There is room to raise the context length or move up a quantisation level.

Generation18.8tok/s
Prompt processing342tok/s
Max context128Ktokens
KV per 1K tokens0GB

Every quantisation of Command R 35B 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 65.2 GB 75.7 GB Runs comfortably 101K 6.73
INT8 / W8A8 8.50 33.7 GB 44.3 GB Runs comfortably 126K 12.2
Q8_0 (GGUF) 8.50 33.7 GB 44.3 GB Runs comfortably 126K 12.2
FP8 (E4M3) 8.00 32.6 GB 43.2 GB Runs comfortably 127K 12.6
Q6_K 6.56 26.7 GB 37.3 GB Runs comfortably 128K 14.9
Q5_K_M 5.67 23.1 GB 33.7 GB Runs comfortably 128K 16.8
AWQ 4-bit 4.25 23.0 GB 33.6 GB Runs comfortably 128K 16.8
GPTQ 4-bit 4.25 23.0 GB 33.6 GB Runs comfortably 128K 16.8
MXFP4 4.25 23.0 GB 33.6 GB Runs comfortably 128K 16.8
Q5_K_S 5.52 22.5 GB 33.1 GB Runs comfortably 128K 17.2
Q4_K_M 4.85 20.1 GB 30.7 GB Runs comfortably 128K 18.8
Q4_K_S 4.58 19.1 GB 29.7 GB Runs comfortably 128K 19.5
Q4_0 4.55 19.0 GB 29.6 GB Runs comfortably 128K 19.6
IQ4_XS 4.25 17.9 GB 28.5 GB Runs comfortably 128K 20.5
Q3_K_M 3.91 16.7 GB 27.3 GB Runs comfortably 128K 21.7
IQ3_M 3.70 15.9 GB 26.5 GB Runs comfortably 128K 22.5
IQ3_XXS 3.06 13.7 GB 24.2 GB Runs comfortably 128K 25.2
Q2_K 2.63 12.1 GB 22.7 GB Runs comfortably 128K 27.5
IQ2_XXS 2.06 10.1 GB 20.7 GB Runs comfortably 128K 31.2
IQ1_M 1.75 9.0 GB 19.5 GB Runs comfortably 128K 33.6

Also worth checking