How much VRAM does Qwen-Image need?
Qwen-Image needs about 20.9 GB at 8-bit, or 12.6 GB as a 4-bit GGUF, at its default output size.
- Type
- Image model
- Parameters
- 20.4B
- Text encoder
- Qwen2.5-VL-7B-Instruct
- Released
- Aug 2025
- License
- Apache 2.0
Memory at each precision
| Precision | Weights | Peak at 1328 × 1328 |
|---|---|---|
| 16-bit BF16 | 38.1 GB | 39.2 GB |
| 8-bit FP8 · Q8_0 | 19.7 GB | 20.9 GB |
| 6-bit GGUF Q6_K | 15.7 GB | 16.8 GB |
| 5-bit GGUF Q5_K_M | 13.6 GB | 14.7 GB |
| 4-bit GGUF Q4_K_M · NVFP4 | 11.4 GB | 12.6 GB |
The Qwen2.5-VL-7B-Instruct text encoder (7.7 GB at FP8) runs first and then waits in system RAM.
What runs it
| Machine | Best fit | Per image |
|---|---|---|
| RTX 3060 12 GB | Too big | |
| RTX 4060 Ti 16 GB | 5-bit | 15 min |
| RTX 3090 | 8-bit | 8.1 min |
| RTX 4090 | 8-bit | 3.5 min |
| RTX 5090 | 8-bit | 2.7 min |
| RTX PRO 6000 Blackwell | 16-bit | 70 s |
| M4 Max 128 GB | 16-bit | 37 min |
| M5 Max 128 GB | 16-bit | 11 min |
| Ryzen AI Max+ 395 128 GB | 16-bit | 17 min |
| DGX Spark 128 GB | 16-bit | 9.2 min |
| M3 Ultra 256 GB | 16-bit | 21 min |
| M5 Ultra 512 GB | 16-bit | 6.6 min |
| H100 SXM | 16-bit | 36 s |
| DGX Station GB300 | 16-bit | 15 s |
1328 × 1328, 50 steps. Times are rough, plus or minus 30%.
Where to get it
Questions
How much VRAM does Qwen-Image need?
Qwen-Image needs about 20.9 GB at 8-bit, or 12.6 GB as a 4-bit GGUF, at its default output size.
Can Qwen-Image run on an RTX 4090?
Yes. At 8-bit it takes about 3.5 min per image.