Skip to content

EXAONE 4.5 33B VRAM requirements

How much VRAM EXAONE 4.5 33B needs
4-bit weights: 19.4 GBWith 8K context: 21.5 GBWith 32K context: 23.0 GB8-bit, 8K context: 36.1 GB
Parameters
34.4B
Longest context
256K
Released
Apr 2026
License
Exaone
Strongest at
Text and images

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

How much VRAM EXAONE 4.5 33B needs

Memory EXAONE 4.5 33B needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit64.0 GB66.1 GB67.6 GB73.6 GB
8-bit34.0 GB36.1 GB37.6 GB43.6 GB
6-bit26.2 GB28.4 GB29.9 GB35.9 GB
5-bit22.8 GB24.9 GB26.4 GB32.4 GB
4-bit19.4 GB21.5 GB23.0 GB29.0 GB
3-bit15.6 GB17.8 GB19.3 GB25.3 GB
2-bit12.0 GB14.1 GB15.6 GB21.6 GB
1.6-bit6.6 GB8.7 GB10.2 GB16.2 GB

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

GPUs and Macs that run EXAONE 4.5 33B

The best way EXAONE 4.5 33B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: EXAONE 4.5 33B on RTX 4060 report
RTX 5060Too bigPDF: EXAONE 4.5 33B on RTX 5060 report
RTX 3070Too bigPDF: EXAONE 4.5 33B on RTX 3070 report
RTX 3080 10 GBToo bigPDF: EXAONE 4.5 33B on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: EXAONE 4.5 33B on RTX 3060 12 GB report
Arc B580Too bigPDF: EXAONE 4.5 33B on Arc B580 report
RTX 4070Too bigPDF: EXAONE 4.5 33B on RTX 4070 report
RTX 4070 SuperToo bigPDF: EXAONE 4.5 33B on RTX 4070 Super report
RTX 5070Too bigPDF: EXAONE 4.5 33B on RTX 5070 report
RTX 4060 Ti 16 GBToo bigPDF: EXAONE 4.5 33B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GBToo bigPDF: EXAONE 4.5 33B on RTX 5060 Ti 16 GB report
Arc A770 16 GBToo bigPDF: EXAONE 4.5 33B on Arc A770 16 GB report
Radeon RX 9070 XTToo bigPDF: EXAONE 4.5 33B on Radeon RX 9070 XT report
RTX 4070 Ti SuperToo bigPDF: EXAONE 4.5 33B on RTX 4070 Ti Super report
RTX 4080 SuperToo bigPDF: EXAONE 4.5 33B on RTX 4080 Super report
RTX 5070 TiToo bigPDF: EXAONE 4.5 33B on RTX 5070 Ti report
RTX 5080Too bigPDF: EXAONE 4.5 33B on RTX 5080 report
Radeon RX 7900 XTX4-bit28K26 tok/s3.2 sPDF: EXAONE 4.5 33B on Radeon RX 7900 XTX report
RTX 30904-bit28K30 tok/s2.1 sPDF: EXAONE 4.5 33B on RTX 3090 report
RTX 40904-bit28K33 tok/s0.9 sPDF: EXAONE 4.5 33B on RTX 4090 report
Radeon AI PRO R97006-bit40K13 tok/s2.1 sPDF: EXAONE 4.5 33B on Radeon AI PRO R9700 report
RTX 50906-bit40K42 tok/s0.7 sPDF: EXAONE 4.5 33B on RTX 5090 report
2× RTX 30908-bit144K18 tok/s1.2 sPDF: EXAONE 4.5 33B on 2× RTX 3090 report
2× RTX 40908-bit144K19 tok/s0.5 sPDF: EXAONE 4.5 33B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit152K25 tok/s1.1 sPDF: EXAONE 4.5 33B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit152K18 tok/s0.8 sPDF: EXAONE 4.5 33B on RTX 6000 Ada report
L40S8-bit152K16 tok/s1.3 sPDF: EXAONE 4.5 33B on L40S report
2× RTX 50908-bit256K33 tok/s0.4 sPDF: EXAONE 4.5 33B on 2× RTX 5090 report
4× RTX 30908-bit256K18 tok/s0.6 sPDF: EXAONE 4.5 33B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit256K33 tok/s0.6 sPDF: EXAONE 4.5 33B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: EXAONE 4.5 33B on Mac mini M4 16 GB report
Mac mini M4 24 GBToo bigPDF: EXAONE 4.5 33B on Mac mini M4 24 GB report
Mac mini M4 32 GB3-bit48K5.8 tok/s47 sPDF: EXAONE 4.5 33B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GBToo bigPDF: EXAONE 4.5 33B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB6-bit104K7.8 tok/s23 sPDF: EXAONE 4.5 33B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit160K6.1 tok/s23 sPDF: EXAONE 4.5 33B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit256K11 tok/s12 sPDF: EXAONE 4.5 33B on M4 Max 128 GB report
M5 Max 128 GB8-bit256K12 tok/s3.4 sPDF: EXAONE 4.5 33B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit256K14 tok/s6.5 sPDF: EXAONE 4.5 33B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit256K14 tok/s6.5 sPDF: EXAONE 4.5 33B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit256K14 tok/s6.5 sPDF: EXAONE 4.5 33B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit256K20 tok/s2.0 sPDF: EXAONE 4.5 33B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit160K4.2 tok/s6.7 sPDF: EXAONE 4.5 33B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit256K4.2 tok/s6.7 sPDF: EXAONE 4.5 33B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit256K5.8 tok/s2.3 sPDF: EXAONE 4.5 33B on RTX Spark report
DGX Spark8-bit256K5.3 tok/s2.3 sPDF: EXAONE 4.5 33B on DGX Spark report
Dual DGX Spark8-bit256K5.3 tok/s1.3 sPDF: EXAONE 4.5 33B on Dual DGX Spark report
DGX Station GB3008-bit256K110 tok/s0.2 sPDF: EXAONE 4.5 33B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit256K33 tok/s0.3 sPDF: EXAONE 4.5 33B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit256K33 tok/s0.2 sPDF: EXAONE 4.5 33B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit256K33 tok/s0.1 sPDF: EXAONE 4.5 33B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit256K37 tok/s1.5 sPDF: EXAONE 4.5 33B on A100 80 GB SXM report
H100 SXM8-bit256K58 tok/s0.5 sPDF: EXAONE 4.5 33B on H100 SXM report
4× H200 NVL8-bit256K80 tok/s0.2 sPDF: EXAONE 4.5 33B on 4× H200 NVL report
8× H200 NVL8-bit256K80 tok/s0.1 sPDF: EXAONE 4.5 33B 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 LGAI-EXAONE/EXAONE-4.5-33B-GGUF
vLLM
vllm serve LGAI-EXAONE/EXAONE-4.5-33B

Questions

How much VRAM does EXAONE 4.5 33B need?

EXAONE 4.5 33B needs 19.4 GB for its weights at 4-bit, 21.5 GB with an 8K context and 23.0 GB with 32K. At 8-bit, 36.1 GB.

What is the minimum VRAM for EXAONE 4.5 33B?

About 17.8 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 Radeon RX 7900 XTX.

How big is EXAONE 4.5 33B?

34.4B parameters: 64.0 GB of weights at 16-bit, 19.4 GB at 4-bit.

Can EXAONE 4.5 33B run on an RTX 4090?

Yes. At 4-bit it holds 28K of context and writes about 33 tokens a second.

More models