Cohere on NVIDIA Hopper

Can I run Command R 35B on an H100 SXM 80GB?

Yes. Command R 35B at Q4_K_M uses 31.0 GB of the 74.4 GB available on a H100 SXM 80GB, and runs at about 91.5 tokens per second. You can push the context to 43K.

Runs comfortably

31.0 GB of 74.4 GB · 42%
074 GB
Weights 20.1 GB
KV cache 10.0 GB
Runtime overhead 0.9 GB

Command R 35B at Q4_K_M leaves 43.4 GB spare on a H100 SXM 80GB. There is room to raise the context length or move up a quantisation level.

Generation91.5tok/s
Prompt processing5937tok/s
Max context43Ktokens
KV per 1K tokens0GB

Every quantisation of Command R 35B on a H100 SXM 80GB

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

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 65.2 GB 76.0 GB Won't fit 7K 9.34
INT8 / W8A8 8.50 33.7 GB 44.6 GB Runs comfortably 32K 59.8
Q8_0 (GGUF) 8.50 33.7 GB 44.6 GB Runs comfortably 32K 59.8
FP8 (E4M3) 8.00 32.6 GB 43.5 GB Runs comfortably 33K 61.5
Q6_K 6.56 26.7 GB 37.6 GB Runs comfortably 37K 72.7
Q5_K_M 5.67 23.1 GB 34.0 GB Runs comfortably 40K 81.9
AWQ 4-bit 4.25 23.0 GB 33.9 GB Runs comfortably 40K 82.0
GPTQ 4-bit 4.25 23.0 GB 33.9 GB Runs comfortably 40K 82.0
MXFP4 4.25 23.0 GB 33.9 GB Runs comfortably 40K 82.0
Q5_K_S 5.52 22.5 GB 33.4 GB Runs comfortably 41K 83.7
Q4_K_M 4.85 20.1 GB 31.0 GB Runs comfortably 43K 91.5
Q4_K_S 4.58 19.1 GB 30.0 GB Runs comfortably 44K 95.1
Q4_0 4.55 19.0 GB 29.9 GB Runs comfortably 44K 95.5
IQ4_XS 4.25 17.9 GB 28.8 GB Runs comfortably 44K 99.9
Q3_K_M 3.91 16.7 GB 27.6 GB Runs comfortably 45K 105
IQ3_M 3.70 15.9 GB 26.8 GB Runs comfortably 46K 109
IQ3_XXS 3.06 13.7 GB 24.5 GB Runs comfortably 48K 122
Q2_K 2.63 12.1 GB 23.0 GB Runs comfortably 49K 133
IQ2_XXS 2.06 10.1 GB 21.0 GB Runs comfortably 51K 150
IQ1_M 1.75 9.0 GB 19.8 GB Runs comfortably 52K 162

Also worth checking

Command R 35B on every GPU
Full hardware tableEverything that runs on a H100 SXM 80GB
Full model tableWhy the KV cache is the part that bites
Guide