Skip to content

EXAONE 3.5 7.8B Instruct VRAM requirements

How much VRAM EXAONE 3.5 7.8B Instruct needs
4-bit weights: 4.4 GBWith 8K context: 6.2 GBWith 32K context: 9.2 GB8-bit, 8K context: 9.5 GB
Parameters
7.8B
Longest context
32K
Released
Dec 2024
License
Exaone
Strongest at
Not rated

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

How much VRAM EXAONE 3.5 7.8B Instruct needs

Memory EXAONE 3.5 7.8B Instruct needs at each precision and context length
PrecisionWeightsWith 8KWith 32K
16-bit14.6 GB16.3 GB19.3 GB
8-bit7.7 GB9.5 GB12.5 GB
6-bit6.0 GB7.7 GB10.7 GB
5-bit5.2 GB6.9 GB9.9 GB
4-bit4.4 GB6.2 GB9.2 GB
3-bit3.6 GB5.3 GB8.3 GB
2-bit2.7 GB4.5 GB7.5 GB
1.6-bit1.5 GB3.3 GB6.3 GB

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

GPUs and Macs that run EXAONE 3.5 7.8B Instruct

The best way EXAONE 3.5 7.8B Instruct runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40604-bit16K41 tok/s1.0 sPDF: EXAONE 3.5 7.8B Instruct on RTX 4060 report
RTX 50604-bit16K66 tok/s0.8 sPDF: EXAONE 3.5 7.8B Instruct on RTX 5060 report
RTX 30704-bit16K66 tok/s0.8 sPDF: EXAONE 3.5 7.8B Instruct on RTX 3070 report
RTX 3080 10 GB6-bit20K81 tok/s0.5 sPDF: EXAONE 3.5 7.8B Instruct on RTX 3080 10 GB report
RTX 3060 12 GB8-bit20K31 tok/s1.2 sPDF: EXAONE 3.5 7.8B Instruct on RTX 3060 12 GB report
Arc B5808-bit20K27 tok/s0.8 sPDF: EXAONE 3.5 7.8B Instruct on Arc B580 report
RTX 40708-bit20K43 tok/s0.5 sPDF: EXAONE 3.5 7.8B Instruct on RTX 4070 report
RTX 4070 Super8-bit20K43 tok/s0.4 sPDF: EXAONE 3.5 7.8B Instruct on RTX 4070 Super report
RTX 50708-bit20K57 tok/s0.5 sPDF: EXAONE 3.5 7.8B Instruct on RTX 5070 report
RTX 4060 Ti 16 GB8-bit32K25 tok/s0.7 sPDF: EXAONE 3.5 7.8B Instruct on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit32K38 tok/s0.7 sPDF: EXAONE 3.5 7.8B Instruct on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit32K33 tok/s0.8 sPDF: EXAONE 3.5 7.8B Instruct on Arc A770 16 GB report
Radeon RX 9070 XT8-bit32K45 tok/s0.4 sPDF: EXAONE 3.5 7.8B Instruct on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit32K57 tok/s0.4 sPDF: EXAONE 3.5 7.8B Instruct on RTX 4070 Ti Super report
RTX 4080 Super8-bit32K62 tok/s0.3 sPDF: EXAONE 3.5 7.8B Instruct on RTX 4080 Super report
RTX 5070 Ti8-bit32K74 tok/s0.4 sPDF: EXAONE 3.5 7.8B Instruct on RTX 5070 Ti report
RTX 50808-bit32K79 tok/s0.3 sPDF: EXAONE 3.5 7.8B Instruct on RTX 5080 report
Radeon RX 7900 XTX8-bit32K66 tok/s0.7 sPDF: EXAONE 3.5 7.8B Instruct on Radeon RX 7900 XTX report
RTX 30908-bit32K77 tok/s0.4 sPDF: EXAONE 3.5 7.8B Instruct on RTX 3090 report
RTX 40908-bit32K83 tok/s0.2 sPDF: EXAONE 3.5 7.8B Instruct on RTX 4090 report
Radeon AI PRO R97008-bit32K45 tok/s0.4 sPDF: EXAONE 3.5 7.8B Instruct on Radeon AI PRO R9700 report
RTX 50908-bit32K138 tok/s0.1 sPDF: EXAONE 3.5 7.8B Instruct on RTX 5090 report
2× RTX 30908-bit32K77 tok/s0.3 sPDF: EXAONE 3.5 7.8B Instruct on 2× RTX 3090 report
2× RTX 40908-bit32K83 tok/s0.1 sPDF: EXAONE 3.5 7.8B Instruct on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit32K107 tok/s0.2 sPDF: EXAONE 3.5 7.8B Instruct on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit32K79 tok/s0.2 sPDF: EXAONE 3.5 7.8B Instruct on RTX 6000 Ada report
L40S8-bit32K72 tok/s0.3 sPDF: EXAONE 3.5 7.8B Instruct on L40S report
2× RTX 50908-bit32K138 tok/s0.1 sPDF: EXAONE 3.5 7.8B Instruct on 2× RTX 5090 report
4× RTX 30908-bit32K77 tok/s0.1 sPDF: EXAONE 3.5 7.8B Instruct on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit32K138 tok/s0.1 sPDF: EXAONE 3.5 7.8B Instruct on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB6-bit28K16 tok/s9.9 sPDF: EXAONE 3.5 7.8B Instruct on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit32K12 tok/s9.9 sPDF: EXAONE 3.5 7.8B Instruct on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit32K12 tok/s9.9 sPDF: EXAONE 3.5 7.8B Instruct on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit32K27 tok/s5.0 sPDF: EXAONE 3.5 7.8B Instruct on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit32K27 tok/s5.0 sPDF: EXAONE 3.5 7.8B Instruct on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit32K27 tok/s5.0 sPDF: EXAONE 3.5 7.8B Instruct on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit32K46 tok/s2.5 sPDF: EXAONE 3.5 7.8B Instruct on M4 Max 128 GB report
M5 Max 128 GB8-bit32K51 tok/s0.7 sPDF: EXAONE 3.5 7.8B Instruct on M5 Max 128 GB report
M3 Ultra 96 GB8-bit32K58 tok/s1.4 sPDF: EXAONE 3.5 7.8B Instruct on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit32K58 tok/s1.4 sPDF: EXAONE 3.5 7.8B Instruct on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit32K58 tok/s1.4 sPDF: EXAONE 3.5 7.8B Instruct on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit32K79 tok/s0.4 sPDF: EXAONE 3.5 7.8B Instruct on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit32K19 tok/s1.4 sPDF: EXAONE 3.5 7.8B Instruct on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit32K19 tok/s1.4 sPDF: EXAONE 3.5 7.8B Instruct on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit32K26 tok/s0.5 sPDF: EXAONE 3.5 7.8B Instruct on RTX Spark report
DGX Spark8-bit32K24 tok/s0.5 sPDF: EXAONE 3.5 7.8B Instruct on DGX Spark report
Dual DGX Spark8-bit32K24 tok/s0.3 sPDF: EXAONE 3.5 7.8B Instruct on Dual DGX Spark report
DGX Station GB3008-bit32K393 tok/s0.1 sPDF: EXAONE 3.5 7.8B Instruct on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit32K138 tok/s0.1 sPDF: EXAONE 3.5 7.8B Instruct on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit32K138 tok/s0.1 sPDF: EXAONE 3.5 7.8B Instruct on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit32K138 tok/s0.1 sPDF: EXAONE 3.5 7.8B Instruct on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit32K155 tok/s0.3 sPDF: EXAONE 3.5 7.8B Instruct on A100 80 GB SXM report
H100 SXM8-bit32K232 tok/s0.1 sPDF: EXAONE 3.5 7.8B Instruct on H100 SXM report
4× H200 NVL8-bit32K303 tok/s0.1 sPDF: EXAONE 3.5 7.8B Instruct on 4× H200 NVL report
8× H200 NVL8-bit32K303 tok/s0.1 sPDF: EXAONE 3.5 7.8B Instruct 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 lmstudio-community/EXAONE-3.5-7.8B-Instruct-GGUF
vLLM
vllm serve LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct
MLX (Mac)
mlx_lm.chat --model mlx-community/EXAONE-3.5-7.8B-Instruct-3bit

Questions

How much VRAM does EXAONE 3.5 7.8B Instruct need?

EXAONE 3.5 7.8B Instruct needs 4.4 GB for its weights at 4-bit, 6.2 GB with an 8K context and 9.2 GB with 32K. At 8-bit, 9.5 GB.

What is the minimum VRAM for EXAONE 3.5 7.8B Instruct?

About 5.3 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 4060.

How big is EXAONE 3.5 7.8B Instruct?

7.8B parameters: 14.6 GB of weights at 16-bit, 4.4 GB at 4-bit.

Can EXAONE 3.5 7.8B Instruct run on an RTX 4090?

Yes. At 8-bit it holds 32K of context and writes about 83 tokens a second.

More models