Skip to content

Glm 4 9B VRAM requirements

How much VRAM Glm 4 9B needs
4-bit weights: 5.3 GBWith 8K context: 10.9 GB8-bit, 8K context: 14.9 GB
Parameters
9.4B
Longest context
8K
Released
Jun 2024
License
GLM 4
Strongest at
Not rated

Its licence (GLM 4) is not a standard open-source one: check the terms before commercial use.

How much VRAM Glm 4 9B needs

Memory Glm 4 9B needs at each precision and context length
PrecisionWeightsWith 8K
16-bit17.5 GB23.1 GB
8-bit9.3 GB14.9 GB
6-bit7.2 GB12.8 GB
5-bit6.2 GB11.8 GB
4-bit5.3 GB10.9 GB
3-bit4.3 GB9.9 GB
2-bit3.3 GB8.9 GB
1.6-bit1.8 GB7.4 GB

One conversation, 16-bit context cache, runtime overhead included.

GPUs and Macs that run Glm 4 9B

The best way Glm 4 9B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Glm 4 9B on RTX 4060 report
RTX 5060Too bigPDF: Glm 4 9B on RTX 5060 report
RTX 3070Too bigPDF: Glm 4 9B on RTX 3070 report
RTX 3080 10 GBToo bigPDF: Glm 4 9B on RTX 3080 10 GB report
RTX 3060 12 GB4-bit8K42 tok/s1.6 sPDF: Glm 4 9B on RTX 3060 12 GB report
Arc B5804-bit8K37 tok/s1.1 sPDF: Glm 4 9B on Arc B580 report
RTX 40704-bit8K58 tok/s0.7 sPDF: Glm 4 9B on RTX 4070 report
RTX 4070 Super4-bit8K58 tok/s0.6 sPDF: Glm 4 9B on RTX 4070 Super report
RTX 50704-bit8K75 tok/s0.7 sPDF: Glm 4 9B on RTX 5070 report
RTX 4060 Ti 16 GB8-bit8K20 tok/s1.0 sPDF: Glm 4 9B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit8K30 tok/s0.9 sPDF: Glm 4 9B on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit8K26 tok/s1.0 sPDF: Glm 4 9B on Arc A770 16 GB report
Radeon RX 9070 XT8-bit8K36 tok/s0.6 sPDF: Glm 4 9B on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit8K45 tok/s0.5 sPDF: Glm 4 9B on RTX 4070 Ti Super report
RTX 4080 Super8-bit8K49 tok/s0.4 sPDF: Glm 4 9B on RTX 4080 Super report
RTX 5070 Ti8-bit8K58 tok/s0.5 sPDF: Glm 4 9B on RTX 5070 Ti report
RTX 50808-bit8K62 tok/s0.4 sPDF: Glm 4 9B on RTX 5080 report
Radeon RX 7900 XTX8-bit8K52 tok/s0.9 sPDF: Glm 4 9B on Radeon RX 7900 XTX report
RTX 30908-bit8K61 tok/s0.6 sPDF: Glm 4 9B on RTX 3090 report
RTX 40908-bit8K65 tok/s0.3 sPDF: Glm 4 9B on RTX 4090 report
Radeon AI PRO R97008-bit8K36 tok/s0.6 sPDF: Glm 4 9B on Radeon AI PRO R9700 report
RTX 50908-bit8K109 tok/s0.2 sPDF: Glm 4 9B on RTX 5090 report
2× RTX 30908-bit8K61 tok/s0.3 sPDF: Glm 4 9B on 2× RTX 3090 report
2× RTX 40908-bit8K65 tok/s0.1 sPDF: Glm 4 9B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit8K85 tok/s0.3 sPDF: Glm 4 9B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit8K62 tok/s0.2 sPDF: Glm 4 9B on RTX 6000 Ada report
L40S8-bit8K56 tok/s0.4 sPDF: Glm 4 9B on L40S report
2× RTX 50908-bit8K109 tok/s0.1 sPDF: Glm 4 9B on 2× RTX 5090 report
4× RTX 30908-bit8K61 tok/s0.2 sPDF: Glm 4 9B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit8K109 tok/s0.2 sPDF: Glm 4 9B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB3-bit8K20 tok/s13 sPDF: Glm 4 9B on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit8K9.6 tok/s13 sPDF: Glm 4 9B on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit8K9.6 tok/s13 sPDF: Glm 4 9B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit8K21 tok/s6.7 sPDF: Glm 4 9B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit8K21 tok/s6.7 sPDF: Glm 4 9B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit8K21 tok/s6.7 sPDF: Glm 4 9B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit8K36 tok/s3.3 sPDF: Glm 4 9B on M4 Max 128 GB report
M5 Max 128 GB8-bit8K40 tok/s1.0 sPDF: Glm 4 9B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit8K46 tok/s1.9 sPDF: Glm 4 9B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit8K46 tok/s1.9 sPDF: Glm 4 9B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit8K46 tok/s1.9 sPDF: Glm 4 9B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit8K62 tok/s0.6 sPDF: Glm 4 9B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit8K15 tok/s1.9 sPDF: Glm 4 9B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit8K15 tok/s1.9 sPDF: Glm 4 9B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit8K21 tok/s0.7 sPDF: Glm 4 9B on RTX Spark report
DGX Spark8-bit8K19 tok/s0.7 sPDF: Glm 4 9B on DGX Spark report
Dual DGX Spark8-bit8K19 tok/s0.4 sPDF: Glm 4 9B on Dual DGX Spark report
DGX Station GB3008-bit8K312 tok/s0.1 sPDF: Glm 4 9B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit8K109 tok/s0.1 sPDF: Glm 4 9B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit8K109 tok/s0.1 sPDF: Glm 4 9B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit8K109 tok/s0.1 sPDF: Glm 4 9B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit8K122 tok/s0.4 sPDF: Glm 4 9B on A100 80 GB SXM report
H100 SXM8-bit8K183 tok/s0.1 sPDF: Glm 4 9B on H100 SXM report
4× H200 NVL8-bit8K240 tok/s0.1 sPDF: Glm 4 9B on 4× H200 NVL report
8× H200 NVL8-bit8K240 tok/s0.1 sPDF: Glm 4 9B on 8× H200 NVL report

Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.

How to run it

llama.cpp
llama-server -hf mradermacher/glm-4-9b-GGUF
vLLM
vllm serve zai-org/glm-4-9b

Questions

How much VRAM does Glm 4 9B need?

Glm 4 9B needs 5.3 GB for its weights at 4-bit, 10.9 GB with an 8K context and 10.9 GB with 8K. At 8-bit, 14.9 GB.

What is the minimum VRAM for Glm 4 9B?

About 9.9 GB at 3-bit with an 8K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the RTX 3060 12 GB.

How big is Glm 4 9B?

9.4B parameters: 17.5 GB of weights at 16-bit, 5.3 GB at 4-bit.

Can Glm 4 9B run on an RTX 4090?

Yes. At 8-bit it holds 8K of context and writes about 65 tokens a second.

More models