Skip to content

Qwen1.5 72B VRAM requirements

How much VRAM Qwen1.5 72B needs
4-bit weights: 40.8 GBWith 8K context: 61.8 GBWith 32K context: 122 GB8-bit, 8K context: 92.5 GB
Parameters
72.3B
Longest context
32K
Released
Jan 2024
License
Tongyi Qianwen
Strongest at
Not rated

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

How much VRAM Qwen1.5 72B needs

Memory Qwen1.5 72B needs at each precision and context length
PrecisionWeightsWith 8KWith 32K
16-bit135 GB156 GB216 GB
8-bit71.5 GB92.5 GB153 GB
6-bit55.2 GB76.2 GB136 GB
5-bit47.9 GB68.9 GB129 GB
4-bit40.8 GB61.8 GB122 GB
3-bit32.9 GB53.9 GB114 GB
2-bit25.2 GB46.2 GB106 GB
1.6-bit13.9 GB34.9 GB94.9 GB

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

GPUs and Macs that run Qwen1.5 72B

The best way Qwen1.5 72B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Qwen1.5 72B on RTX 4060 report
RTX 5060Too bigPDF: Qwen1.5 72B on RTX 5060 report
RTX 3070Too bigPDF: Qwen1.5 72B on RTX 3070 report
RTX 3080 10 GBToo bigPDF: Qwen1.5 72B on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: Qwen1.5 72B on RTX 3060 12 GB report
Arc B580Too bigPDF: Qwen1.5 72B on Arc B580 report
RTX 4070Too bigPDF: Qwen1.5 72B on RTX 4070 report
RTX 4070 SuperToo bigPDF: Qwen1.5 72B on RTX 4070 Super report
RTX 5070Too bigPDF: Qwen1.5 72B on RTX 5070 report
RTX 4060 Ti 16 GBToo bigPDF: Qwen1.5 72B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GBToo bigPDF: Qwen1.5 72B on RTX 5060 Ti 16 GB report
Arc A770 16 GBToo bigPDF: Qwen1.5 72B on Arc A770 16 GB report
Radeon RX 9070 XTToo bigPDF: Qwen1.5 72B on Radeon RX 9070 XT report
RTX 4070 Ti SuperToo bigPDF: Qwen1.5 72B on RTX 4070 Ti Super report
RTX 4080 SuperToo bigPDF: Qwen1.5 72B on RTX 4080 Super report
RTX 5070 TiToo bigPDF: Qwen1.5 72B on RTX 5070 Ti report
RTX 5080Too bigPDF: Qwen1.5 72B on RTX 5080 report
Radeon RX 7900 XTXToo bigPDF: Qwen1.5 72B on Radeon RX 7900 XTX report
RTX 3090Too bigPDF: Qwen1.5 72B on RTX 3090 report
RTX 4090Too bigPDF: Qwen1.5 72B on RTX 4090 report
Radeon AI PRO R9700Too bigPDF: Qwen1.5 72B on Radeon AI PRO R9700 report
RTX 5090Too bigPDF: Qwen1.5 72B on RTX 5090 report
2× RTX 3090Too bigPDF: Qwen1.5 72B on 2× RTX 3090 report
2× RTX 4090Too bigPDF: Qwen1.5 72B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GBToo bigPDF: Qwen1.5 72B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 AdaToo bigPDF: Qwen1.5 72B on RTX 6000 Ada report
L40SToo bigPDF: Qwen1.5 72B on L40S report
2× RTX 5090Too bigPDF: Qwen1.5 72B on 2× RTX 5090 report
4× RTX 30905-bit16K13 tok/s1.3 sPDF: Qwen1.5 72B on 4× RTX 3090 report
RTX PRO 6000 Blackwell5-bit16K23 tok/s1.2 sPDF: Qwen1.5 72B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: Qwen1.5 72B on Mac mini M4 16 GB report
Mac mini M4 24 GBToo bigPDF: Qwen1.5 72B on Mac mini M4 24 GB report
Mac mini M4 32 GBToo bigPDF: Qwen1.5 72B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GBToo bigPDF: Qwen1.5 72B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GBToo bigPDF: Qwen1.5 72B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GBToo bigPDF: Qwen1.5 72B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB5-bit16K7.7 tok/s24 sPDF: Qwen1.5 72B on M4 Max 128 GB report
M5 Max 128 GB5-bit16K8.6 tok/s7.0 sPDF: Qwen1.5 72B on M5 Max 128 GB report
M3 Ultra 96 GBToo bigPDF: Qwen1.5 72B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit32K6.7 tok/s14 sPDF: Qwen1.5 72B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit32K6.7 tok/s14 sPDF: Qwen1.5 72B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit32K9.7 tok/s4.3 sPDF: Qwen1.5 72B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GBToo bigPDF: Qwen1.5 72B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB5-bit16K3.0 tok/s14 sPDF: Qwen1.5 72B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit16K2.8 tok/s4.9 sPDF: Qwen1.5 72B on RTX Spark report
DGX Spark8-bit16K2.5 tok/s4.9 sPDF: Qwen1.5 72B on DGX Spark report
Dual DGX Spark8-bit32K2.5 tok/s2.8 sPDF: Qwen1.5 72B on Dual DGX Spark report
DGX Station GB3008-bit32K57 tok/s0.4 sPDF: Qwen1.5 72B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit32K16 tok/s0.7 sPDF: Qwen1.5 72B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit32K16 tok/s0.4 sPDF: Qwen1.5 72B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit32K16 tok/s0.2 sPDF: Qwen1.5 72B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM3-bit16K37 tok/s3.1 sPDF: Qwen1.5 72B on A100 80 GB SXM report
H100 SXM3-bit16K58 tok/s1.0 sPDF: Qwen1.5 72B on H100 SXM report
4× H200 NVL8-bit32K40 tok/s0.4 sPDF: Qwen1.5 72B on 4× H200 NVL report
8× H200 NVL8-bit32K40 tok/s0.2 sPDF: Qwen1.5 72B 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/Qwen1.5-72B-i1-GGUF
vLLM
vllm serve Qwen/Qwen1.5-72B

Questions

How much VRAM does Qwen1.5 72B need?

Qwen1.5 72B needs 40.8 GB for its weights at 4-bit, 61.8 GB with an 8K context and 122 GB with 32K. At 8-bit, 92.5 GB.

What is the minimum VRAM for Qwen1.5 72B?

About 53.9 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 4× RTX 3090.

How big is Qwen1.5 72B?

72.3B parameters: 135 GB of weights at 16-bit, 40.8 GB at 4-bit.

Can Qwen1.5 72B run on an RTX 4090?

Not on one card. The smallest machine here that runs it is the 4× RTX 3090.

More models