Zhipu · GLM-4 · 9.4B parameters

GLM-4 9B VRAM requirements

GLM-4 9B has 40 layers and uses grouped-query attention (2 KV heads). At Q4_K_M the weights come to 5.4 GB, and the best quantisation that fits a 24 GB card is FP16 / BF16.

Model

Hardware

Quantisation

Context 8K

KV cache precision

Runtime

Batch

GPUs

Monitor plugged into this GPU

Runs comfortably

6.5 GB of 21.8 GB · 30%

022 GB

Weights 5.4 GB

KV cache 0.3 GB

Runtime overhead 0.8 GB

GLM-4 9B at Q4_K_M leaves 15.2 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.

Generation106tok/s

Prompt processing3686tok/s

Max context128Ktokens

KV per 1K tokens0GB

Every quantisation of GLM-4 9B on a RTX 4090

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

Quantisation	bpw	Weights	Total	Verdict	Max ctx	tok/s
FP16 / BF16	16.00	17.5 GB	18.6 GB	Runs comfortably	88K	33.9
INT8 / W8A8	8.50	9.0 GB	10.2 GB	Runs comfortably	128K	64.8
Q8_0 (GGUF)	8.50	9.0 GB	10.2 GB	Runs comfortably	128K	64.8
FP8 (E4M3)	8.00	8.8 GB	9.9 GB	Runs comfortably	128K	66.8
Q6_K	6.56	7.2 GB	8.3 GB	Runs comfortably	128K	80.8
AWQ 4-bit	4.25	6.3 GB	7.5 GB	Runs comfortably	128K	90.9
GPTQ 4-bit	4.25	6.3 GB	7.5 GB	Runs comfortably	128K	90.9
MXFP4	4.25	6.3 GB	7.5 GB	Runs comfortably	128K	90.9
Q5_K_M	5.67	6.2 GB	7.3 GB	Runs comfortably	128K	92.9
Q5_K_S	5.52	6.0 GB	7.2 GB	Runs comfortably	128K	95.3
Q4_K_M	4.85	5.4 GB	6.5 GB	Runs comfortably	128K	106
Q4_K_S	4.58	5.1 GB	6.3 GB	Runs comfortably	128K	111
Q4_0	4.55	5.1 GB	6.2 GB	Runs comfortably	128K	112
IQ4_XS	4.25	4.8 GB	6.0 GB	Runs comfortably	128K	118
Q3_K_M	3.91	4.5 GB	5.6 GB	Runs comfortably	128K	126
IQ3_M	3.70	4.3 GB	5.4 GB	Runs comfortably	128K	131
IQ3_XXS	3.06	3.7 GB	4.8 GB	Runs comfortably	128K	151
Q2_K	2.63	3.3 GB	4.4 GB	Runs comfortably	128K	168
IQ2_XXS	2.06	2.8 GB	3.9 GB	Runs comfortably	128K	198
IQ1_M	1.75	2.5 GB	3.6 GB	Runs comfortably	128K	219

GLM-4 9B on each GPU

Q4_K_M weights at 8K context, single card, monitor attached.

GPU	VRAM	GB/s	Verdict	Max ctx	tok/s
H100 SXM 80GB	80	3350	Runs comfortably	128K	383
A100 80GB	80	2039	Runs comfortably	128K	218
RTX 5090	32	1792	Runs comfortably	128K	203
RTX 5080	16	960	Runs comfortably	128K	111
RTX 4090	24	1008	Runs comfortably	128K	106
RTX 5070 Ti	16	896	Runs comfortably	128K	104
RTX 3090	24	936	Runs comfortably	128K	103
Radeon RX 7900 XTX	24	960	Runs comfortably	128K	96.0
L40S	48	864	Runs comfortably	128K	91.2
RTX A6000	48	768	Runs comfortably	128K	84.7
RTX 3080 10GB	10	760	Runs comfortably	61K	83.9
Mac Studio M3 Ultra 256GB	256	819	Runs comfortably	128K	83.4
RTX 5070	12	672	Runs comfortably	109K	78.4
RTX 4080 Super	16	736	Runs comfortably	128K	77.9
RTX 4070 Ti Super	16	672	Runs comfortably	128K	71.3
Mac Studio M4 Max 128GB	128	546	Runs comfortably	128K	62.3
RTX 4070 Super	12	504	Runs comfortably	109K	53.7
RTX 4070	12	504	Runs comfortably	109K	53.7
RTX 5060 Ti 16GB	16	448	Runs comfortably	128K	52.6
Arc B580	12	456	Runs comfortably	109K	41.2
RTX 3060 12GB	12	360	Runs comfortably	109K	40.2
NVIDIA DGX Spark (GB10)	128	273	Runs comfortably	128K	32.2
Mac Mini M4 Pro 48GB	48	273	Runs comfortably	128K	31.4
RTX 4060 Ti 16GB	16	288	Runs comfortably	128K	30.8
Ryzen AI Max+ 395 128GB	128	256	Runs comfortably	128K	24.1

Architecture

Parameters	9.4B
Layers	40
Hidden size	4096
Attention heads / KV heads	32 / 2
Head dimension	128
Vocabulary	151,552
Trained context	128K
KV cache per 1K tokens	0 GB
Hugging Face	THUDM/glm-4-9b-chat

The Zhipu family

GLM-4 9B is published on Hugging Face under THUDM, the Tsinghua lab that develops the series with Zhipu AI. A 9.4B dense model with a 128K context.

huggingface.co/THUDM · github.com/THUDM/GLM-4

Direct answers

GLM-4 9B on RTX 5090
See the verdict GLM-4 9B on RTX 4090
See the verdict GLM-4 9B on RTX 3090
See the verdict GLM-4 9B on RTX 5080
See the verdict GLM-4 9B on RTX 5070 Ti
See the verdict GLM-4 9B on RTX 5070
See the verdict GLM-4 9B on RTX 4070 Ti Super
See the verdict GLM-4 9B on RTX 4070 Super
See the verdict GLM-4 9B on RTX 4070
See the verdict