Cohere on NVIDIA Ada

Can I run Command R 35B on an L40S?

Yes. Command R 35B at Q4_K_M uses 31.0 GB of the 44.3 GB available on a L40S, and runs at about 20.6 tokens per second. You can push the context to 19K.

Runs comfortably

31.0 GB of 44.3 GB · 70%
044 GB
Weights 20.1 GB
KV cache 10.0 GB
Runtime overhead 0.9 GB

Command R 35B at Q4_K_M leaves 13.4 GB spare on a L40S. There is room to raise the context length or move up a quantisation level.

Generation20.6tok/s
Prompt processing1087tok/s
Max context19Ktokens
KV per 1K tokens0GB

Every quantisation of Command R 35B on a L40S

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 — 1.02
INT8 / W8A8 8.50 33.7 GB 44.6 GB Won't fit 8K 10.2
Q8_0 (GGUF) 8.50 33.7 GB 44.6 GB Won't fit 8K 10.2
FP8 (E4M3) 8.00 32.6 GB 43.5 GB Fits, but tight 9K 13.7
Q6_K 6.56 26.7 GB 37.6 GB Runs comfortably 13K 16.3
Q5_K_M 5.67 23.1 GB 34.0 GB Runs comfortably 16K 18.4
AWQ 4-bit 4.25 23.0 GB 33.9 GB Runs comfortably 16K 18.4
GPTQ 4-bit 4.25 23.0 GB 33.9 GB Runs comfortably 16K 18.4
MXFP4 4.25 23.0 GB 33.9 GB Runs comfortably 16K 18.4
Q5_K_S 5.52 22.5 GB 33.4 GB Runs comfortably 17K 18.8
Q4_K_M 4.85 20.1 GB 31.0 GB Runs comfortably 19K 20.6
Q4_K_S 4.58 19.1 GB 30.0 GB Runs comfortably 19K 21.4
Q4_0 4.55 19.0 GB 29.9 GB Runs comfortably 20K 21.5
IQ4_XS 4.25 17.9 GB 28.8 GB Runs comfortably 20K 22.5
Q3_K_M 3.91 16.7 GB 27.6 GB Runs comfortably 21K 23.7
IQ3_M 3.70 15.9 GB 26.8 GB Runs comfortably 22K 24.6
IQ3_XXS 3.06 13.7 GB 24.5 GB Runs comfortably 24K 27.6
Q2_K 2.63 12.1 GB 23.0 GB Runs comfortably 25K 30.1
IQ2_XXS 2.06 10.1 GB 21.0 GB Runs comfortably 27K 34.1
IQ1_M 1.75 9.0 GB 19.8 GB Runs comfortably 28K 36.8

Also worth checking

Command R 35B on every GPU
Full hardware tableEverything that runs on a L40S
Full model tableWhy the KV cache is the part that bites
Guide