Skip to content

WizardLM 2 8x22B VRAM requirements

How much VRAM WizardLM 2 8x22B needs
4-bit weights: 79.4 GBWith 8K context: 81.8 GBWith 32K context: 87.1 GB8-bit, 8K context: 142 GB
Parameters
141B
Active per token
39.2B
Longest context
64K
Released
Apr 2024
License
Apache 2.0
Strongest at
Not rated

How much VRAM WizardLM 2 8x22B needs

Memory WizardLM 2 8x22B needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 64K
16-bit262 GB264 GB270 GB277 GB
8-bit139 GB142 GB147 GB154 GB
6-bit107 GB110 GB115 GB122 GB
5-bit93.1 GB95.6 GB101 GB108 GB
4-bit79.4 GB81.8 GB87.1 GB94.1 GB
3-bit64.0 GB66.5 GB71.7 GB78.7 GB
2-bit49.1 GB51.6 GB56.8 GB63.8 GB
1.6-bit27.0 GB29.5 GB34.7 GB41.7 GB

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

GPUs and Macs that run WizardLM 2 8x22B

The best way WizardLM 2 8x22B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: WizardLM 2 8x22B on RTX 4060 report
RTX 5060Too bigPDF: WizardLM 2 8x22B on RTX 5060 report
RTX 3070Too bigPDF: WizardLM 2 8x22B on RTX 3070 report
RTX 3080 10 GBToo bigPDF: WizardLM 2 8x22B on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: WizardLM 2 8x22B on RTX 3060 12 GB report
Arc B580Too bigPDF: WizardLM 2 8x22B on Arc B580 report
RTX 4070Too bigPDF: WizardLM 2 8x22B on RTX 4070 report
RTX 4070 SuperToo bigPDF: WizardLM 2 8x22B on RTX 4070 Super report
RTX 5070Too bigPDF: WizardLM 2 8x22B on RTX 5070 report
RTX 4060 Ti 16 GBToo bigPDF: WizardLM 2 8x22B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GBToo bigPDF: WizardLM 2 8x22B on RTX 5060 Ti 16 GB report
Arc A770 16 GBToo bigPDF: WizardLM 2 8x22B on Arc A770 16 GB report
Radeon RX 9070 XTToo bigPDF: WizardLM 2 8x22B on Radeon RX 9070 XT report
RTX 4070 Ti SuperToo bigPDF: WizardLM 2 8x22B on RTX 4070 Ti Super report
RTX 4080 SuperToo bigPDF: WizardLM 2 8x22B on RTX 4080 Super report
RTX 5070 TiToo bigPDF: WizardLM 2 8x22B on RTX 5070 Ti report
RTX 5080Too bigPDF: WizardLM 2 8x22B on RTX 5080 report
Radeon RX 7900 XTXToo bigPDF: WizardLM 2 8x22B on Radeon RX 7900 XTX report
RTX 3090Too bigPDF: WizardLM 2 8x22B on RTX 3090 report
RTX 4090Too bigPDF: WizardLM 2 8x22B on RTX 4090 report
Radeon AI PRO R9700Too bigPDF: WizardLM 2 8x22B on Radeon AI PRO R9700 report
RTX 5090Too bigPDF: WizardLM 2 8x22B on RTX 5090 report
2× RTX 3090Too bigPDF: WizardLM 2 8x22B on 2× RTX 3090 report
2× RTX 4090Too bigPDF: WizardLM 2 8x22B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GBToo bigPDF: WizardLM 2 8x22B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 AdaToo bigPDF: WizardLM 2 8x22B on RTX 6000 Ada report
L40SToo bigPDF: WizardLM 2 8x22B on L40S report
2× RTX 50902-bit44K66 tok/s1.5 sPDF: WizardLM 2 8x22B on 2× RTX 5090 report
4× RTX 30904-bit40K25 tok/s2.5 sPDF: WizardLM 2 8x22B on 4× RTX 3090 report
RTX PRO 6000 Blackwell4-bit48K45 tok/s2.2 sPDF: WizardLM 2 8x22B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: WizardLM 2 8x22B on Mac mini M4 16 GB report
Mac mini M4 24 GBToo bigPDF: WizardLM 2 8x22B on Mac mini M4 24 GB report
Mac mini M4 32 GBToo bigPDF: WizardLM 2 8x22B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GBToo bigPDF: WizardLM 2 8x22B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GBToo bigPDF: WizardLM 2 8x22B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GBToo bigPDF: WizardLM 2 8x22B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB4-bit48K15 tok/s45 sPDF: WizardLM 2 8x22B on M4 Max 128 GB report
M5 Max 128 GB4-bit48K17 tok/s13 sPDF: WizardLM 2 8x22B on M5 Max 128 GB report
M3 Ultra 96 GB3-bit16K23 tok/s25 sPDF: WizardLM 2 8x22B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit64K12 tok/s25 sPDF: WizardLM 2 8x22B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit64K12 tok/s25 sPDF: WizardLM 2 8x22B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit64K16 tok/s7.9 sPDF: WizardLM 2 8x22B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GBToo bigPDF: WizardLM 2 8x22B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB4-bit48K6.3 tok/s26 sPDF: WizardLM 2 8x22B on Ryzen AI Max+ 395 128 GB report
RTX Spark6-bit28K6.4 tok/s9.1 sPDF: WizardLM 2 8x22B on RTX Spark report
DGX Spark6-bit28K5.8 tok/s9.1 sPDF: WizardLM 2 8x22B on DGX Spark report
Dual DGX Spark8-bit64K4.5 tok/s5.2 sPDF: WizardLM 2 8x22B on Dual DGX Spark report
DGX Station GB3008-bit64K85 tok/s0.7 sPDF: WizardLM 2 8x22B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit64K27 tok/s1.2 sPDF: WizardLM 2 8x22B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit64K27 tok/s0.7 sPDF: WizardLM 2 8x22B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit64K27 tok/s0.3 sPDF: WizardLM 2 8x22B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM3-bit48K60 tok/s5.8 sPDF: WizardLM 2 8x22B on A100 80 GB SXM report
H100 SXM3-bit48K87 tok/s1.8 sPDF: WizardLM 2 8x22B on H100 SXM report
4× H200 NVL8-bit64K64 tok/s0.7 sPDF: WizardLM 2 8x22B on 4× H200 NVL report
8× H200 NVL8-bit64K64 tok/s0.3 sPDF: WizardLM 2 8x22B 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 qwp4w3hyb/Not-WizardLM-2-8x22B-iMat-GGUF
vLLM
vllm serve alpindale/WizardLM-2-8x22B

Questions

How much VRAM does WizardLM 2 8x22B need?

WizardLM 2 8x22B needs 79.4 GB for its weights at 4-bit, 81.8 GB with an 8K context and 87.1 GB with 32K. At 8-bit, 142 GB.

What is the minimum VRAM for WizardLM 2 8x22B?

About 51.6 GB at 2-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 2× RTX 5090.

How big is WizardLM 2 8x22B?

141B parameters (39.2B active per token): 262 GB of weights at 16-bit, 79.4 GB at 4-bit.

Can WizardLM 2 8x22B run on an RTX 4090?

Not on one card. The smallest machine here that runs it is the 2× RTX 5090.

More models