Qwen-Image VRAM Requirements
Qwen-Image needs about 20.4 GB of VRAM with FP8 weights at 1328×1328, loading one model at a time. In FP16 it needs 39.4 GB (59.8 GB with everything loaded), and GGUF Q4 brings it to 12.3 GB. The smallest consumer GPU that runs it is the RTX 4060 Ti 16GB with Q4 weights, or the RTX 3090 24GB in FP8.
20.4 GB peak VRAM
Qwen-Image at 1328×1328 with FP8 weights, loading one model at a time (ComfyUI default on smaller cards). Keeping everything loaded needs 33.1 GB.
- Diffusion model 19.0 GB
- Text encoders 7.7 GB
- VAE 0.2 GB
- Peak activations 5.3 GB
- Overhead 0.8 GB
GPU fit
- RTX 4060 8GB37% on GPU
- RTX 3060 12GB55% on GPU
- RTX 4060 Ti 16GB74% on GPU
- RTX 4090 24GBFitswith swapping
- RTX 5090 32GBFitswith swapping
- RTX 6000 Ada 48GBFitsall loaded
- RTX PRO 6000 Blackwell 96GBFitsall loaded
- Apple M4 Max 128GBFitsall loaded
The VAE decode is the peak here. Turn on tiled VAE decoding to cut it to about 1.3 GB.
Qwen-Image VRAM by precision
At 1328×1328. Text encoders in FP8 except for the FP16 row.
| Weights | Diffusion model | Peak (swapping) | All loaded | Smallest tier |
|---|---|---|---|---|
| FP16 / BF16 | 38.0 GB | 39.4 GB | 59.8 GB | 48 GB |
| FP8 | 19.0 GB | 20.4 GB | 33.1 GB | 24 GB |
| GGUF Q8_0 | 20.2 GB | 21.6 GB | 34.2 GB | 24 GB |
| GGUF Q5_K_S | 13.1 GB | 14.5 GB | 27.2 GB | 16 GB |
| GGUF Q4_K_S / NF4 | 10.8 GB | 12.3 GB | 24.9 GB | 16 GB |
VRAM by resolution
FP8 weights, with model swapping.
| Resolution | Latent tokens | Activations | Peak VRAM |
|---|---|---|---|
| 1024×1024 | 4,096 | 0.4 GB | 20.2 GB |
| 1536×1536 | 9,216 | 0.8 GB | 20.6 GB |
About Qwen-Image
A 20B image model known for rendering text well, including Chinese. Its text encoder is a full 7B vision-language model, so the encoder alone takes about 15 GB in BF16. FP8 or GGUF weights plus model swapping are required on consumer cards.
| Component | Size |
|---|---|
| Diffusion model | 20.4B parameters (MMDiT) |
| Text encoders | Qwen2.5-VL 7B (8.3B) |
| VAE | 0.127B parameters |
| Classifier-free guidance | Yes (2 passes per step) |
| Released | 2025-08 |
| License | Apache 2.0 |
Qwen-Image VRAM: common questions
Can Qwen-Image run on a 24GB GPU?
Yes with FP8 weights (about 19 GB) and the text encoder swapped out during sampling, or GGUF Q4 to Q5 for headroom.
How much VRAM does Qwen-Image need?
At its default 1328×1328: about 39.4 GB in FP16 with model swapping (59.8 GB with everything loaded), 20.4 GB in FP8 and 12.3 GB with GGUF Q4 weights.