Skip to content

Qwen Image Bench VRAM requirements

How much VRAM Qwen Image Bench needs
4-bit weights: 15.4 GBWith 8K context: 17.2 GBWith 32K context: 18.7 GB8-bit, 8K context: 28.8 GB
Parameters
27.4B
Longest context
256K
Released
May 2026
License
Apache 2.0
Strongest at
Text and images

How much VRAM Qwen Image Bench needs

Memory Qwen Image Bench needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit51.0 GB52.7 GB54.2 GB60.2 GB
8-bit27.1 GB28.8 GB30.3 GB36.3 GB
6-bit20.9 GB22.6 GB24.1 GB30.1 GB
5-bit18.1 GB19.8 GB21.3 GB27.3 GB
4-bit15.4 GB17.2 GB18.7 GB24.7 GB
3-bit12.5 GB14.2 GB15.7 GB21.7 GB
2-bit9.6 GB11.3 GB12.8 GB18.8 GB
1.6-bit5.3 GB7.0 GB8.5 GB14.5 GB

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

GPUs and Macs that run Qwen Image Bench

The best way Qwen Image Bench runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Qwen Image Bench on RTX 4060 report
RTX 5060Too bigPDF: Qwen Image Bench on RTX 5060 report
RTX 3070Too bigPDF: Qwen Image Bench on RTX 3070 report
RTX 3080 10 GBToo bigPDF: Qwen Image Bench on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: Qwen Image Bench on RTX 3060 12 GB report
Arc B580Too bigPDF: Qwen Image Bench on Arc B580 report
RTX 4070Too bigPDF: Qwen Image Bench on RTX 4070 report
RTX 4070 SuperToo bigPDF: Qwen Image Bench on RTX 4070 Super report
RTX 5070Too bigPDF: Qwen Image Bench on RTX 5070 report
RTX 4060 Ti 16 GB3-bit24K15 tok/s2.5 sPDF: Qwen Image Bench on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB3-bit24K23 tok/s2.3 sPDF: Qwen Image Bench on RTX 5060 Ti 16 GB report
Arc A770 16 GB3-bit24K20 tok/s2.7 sPDF: Qwen Image Bench on Arc A770 16 GB report
Radeon RX 9070 XT3-bit24K27 tok/s1.5 sPDF: Qwen Image Bench on Radeon RX 9070 XT report
RTX 4070 Ti Super3-bit24K34 tok/s1.2 sPDF: Qwen Image Bench on RTX 4070 Ti Super report
RTX 4080 Super3-bit24K37 tok/s1.1 sPDF: Qwen Image Bench on RTX 4080 Super report
RTX 5070 Ti3-bit24K44 tok/s1.3 sPDF: Qwen Image Bench on RTX 5070 Ti report
RTX 50803-bit24K47 tok/s1.0 sPDF: Qwen Image Bench on RTX 5080 report
Radeon RX 7900 XTX5-bit52K28 tok/s2.4 sPDF: Qwen Image Bench on Radeon RX 7900 XTX report
RTX 30905-bit52K33 tok/s1.5 sPDF: Qwen Image Bench on RTX 3090 report
RTX 40905-bit52K35 tok/s0.7 sPDF: Qwen Image Bench on RTX 4090 report
Radeon AI PRO R97008-bit32K13 tok/s1.6 sPDF: Qwen Image Bench on Radeon AI PRO R9700 report
RTX 50908-bit32K41 tok/s0.5 sPDF: Qwen Image Bench on RTX 5090 report
2× RTX 30908-bit256K22 tok/s0.9 sPDF: Qwen Image Bench on 2× RTX 3090 report
2× RTX 40908-bit256K24 tok/s0.4 sPDF: Qwen Image Bench on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit256K32 tok/s0.8 sPDF: Qwen Image Bench on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit256K23 tok/s0.6 sPDF: Qwen Image Bench on RTX 6000 Ada report
L40S8-bit256K21 tok/s1.0 sPDF: Qwen Image Bench on L40S report
2× RTX 50908-bit256K41 tok/s0.3 sPDF: Qwen Image Bench on 2× RTX 5090 report
4× RTX 30908-bit256K22 tok/s0.5 sPDF: Qwen Image Bench on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit256K41 tok/s0.4 sPDF: Qwen Image Bench on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: Qwen Image Bench on Mac mini M4 16 GB report
Mac mini M4 24 GB3-bit24K7.3 tok/s35 sPDF: Qwen Image Bench on Mac mini M4 24 GB report
Mac mini M4 32 GB5-bit16K5.1 tok/s35 sPDF: Qwen Image Bench on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB3-bit24K16 tok/s17 sPDF: Qwen Image Bench on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit96K7.7 tok/s17 sPDF: Qwen Image Bench on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit256K7.7 tok/s17 sPDF: Qwen Image Bench on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit256K14 tok/s8.7 sPDF: Qwen Image Bench on M4 Max 128 GB report
M5 Max 128 GB8-bit256K15 tok/s2.5 sPDF: Qwen Image Bench on M5 Max 128 GB report
M3 Ultra 96 GB8-bit256K17 tok/s4.9 sPDF: Qwen Image Bench on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit256K17 tok/s4.9 sPDF: Qwen Image Bench on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit256K17 tok/s4.9 sPDF: Qwen Image Bench on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit256K24 tok/s1.5 sPDF: Qwen Image Bench on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit256K5.4 tok/s5.0 sPDF: Qwen Image Bench on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit256K5.4 tok/s5.0 sPDF: Qwen Image Bench on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit256K7.4 tok/s1.8 sPDF: Qwen Image Bench on RTX Spark report
DGX Spark8-bit256K6.8 tok/s1.8 sPDF: Qwen Image Bench on DGX Spark report
Dual DGX Spark8-bit256K6.8 tok/s1.0 sPDF: Qwen Image Bench on Dual DGX Spark report
DGX Station GB3008-bit256K133 tok/s0.1 sPDF: Qwen Image Bench on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit256K41 tok/s0.2 sPDF: Qwen Image Bench on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit256K41 tok/s0.1 sPDF: Qwen Image Bench on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit256K41 tok/s0.1 sPDF: Qwen Image Bench on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit256K47 tok/s1.1 sPDF: Qwen Image Bench on A100 80 GB SXM report
H100 SXM8-bit256K72 tok/s0.4 sPDF: Qwen Image Bench on H100 SXM report
4× H200 NVL8-bit256K98 tok/s0.1 sPDF: Qwen Image Bench on 4× H200 NVL report
8× H200 NVL8-bit256K98 tok/s0.1 sPDF: Qwen Image Bench 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 DawnW0lf/Qwen-Image-Bench-GGUF
vLLM
vllm serve Qwen/Qwen-Image-Bench

Questions

How much VRAM does Qwen Image Bench need?

Qwen Image Bench needs 15.4 GB for its weights at 4-bit, 17.2 GB with an 8K context and 18.7 GB with 32K. At 8-bit, 28.8 GB.

What is the minimum VRAM for Qwen Image Bench?

About 14.2 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 Ti 16 GB.

How big is Qwen Image Bench?

27.4B parameters: 51.0 GB of weights at 16-bit, 15.4 GB at 4-bit.

Can Qwen Image Bench run on an RTX 4090?

Yes. At 5-bit it holds 52K of context and writes about 35 tokens a second.

More models