Zhipu on NVIDIA Ampere

Can I run GLM-4 9B on an RTX 3080 10GB?

Yes. GLM-4 9B at Q4_K_M uses 6.5 GB of the 8.6 GB available on a RTX 3080 10GB, and runs at about 83.9 tokens per second. You can push the context to 61K.

Runs comfortably

6.5 GB of 8.6 GB · 76%
09 GB
Weights 5.4 GB
KV cache 0.3 GB
Runtime overhead 0.8 GB

GLM-4 9B at Q4_K_M leaves 2.1 GB spare on a RTX 3080 10GB. There is room to raise the context length or move up a quantisation level.

Generation83.9tok/s
Prompt processing1340tok/s
Max context61Ktokens
KV per 1K tokens0GB

Every quantisation of GLM-4 9B on a RTX 3080 10GB

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

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 17.5 GB 18.6 GB Won't fit 3.70
INT8 / W8A8 8.50 9.0 GB 10.2 GB Won't fit 17.7
Q8_0 (GGUF) 8.50 9.0 GB 10.2 GB Won't fit 17.7
FP8 (E4M3) 8.00 8.8 GB 9.9 GB Won't fit 20.2
Q6_K 6.56 7.2 GB 8.3 GB Fits, but tight 15K 63.9
AWQ 4-bit 4.25 6.3 GB 7.5 GB Runs comfortably 37K 71.9
GPTQ 4-bit 4.25 6.3 GB 7.5 GB Runs comfortably 37K 71.9
MXFP4 4.25 6.3 GB 7.5 GB Runs comfortably 37K 71.9
Q5_K_M 5.67 6.2 GB 7.3 GB Runs comfortably 40K 73.5
Q5_K_S 5.52 6.0 GB 7.2 GB Runs comfortably 45K 75.4
Q4_K_M 4.85 5.4 GB 6.5 GB Runs comfortably 61K 83.9
Q4_K_S 4.58 5.1 GB 6.3 GB Runs comfortably 67K 87.8
Q4_0 4.55 5.1 GB 6.2 GB Runs comfortably 68K 88.3
IQ4_XS 4.25 4.8 GB 6.0 GB Runs comfortably 75K 93.2
Q3_K_M 3.91 4.5 GB 5.6 GB Runs comfortably 84K 99.5
IQ3_M 3.70 4.3 GB 5.4 GB Runs comfortably 89K 104
IQ3_XXS 3.06 3.7 GB 4.8 GB Runs comfortably 104K 120
Q2_K 2.63 3.3 GB 4.4 GB Runs comfortably 115K 134
IQ2_XXS 2.06 2.8 GB 3.9 GB Runs comfortably 128K 157
IQ1_M 1.75 2.5 GB 3.6 GB Runs comfortably 128K 174

Also worth checking