Qwen-Image VRAM requirements: how big are the model files?

Short answer: the Diffusers files for Qwen-Image add up to 57.7 GB on Hugging Face, and the image transformer alone is 40.9 GB in BF16.^[1] That is weights only. It does not include the memory used while an image is generated, which depends on resolution and settings. Qwen does not publish an official VRAM figure on the model card, so we don't invent one.^[2]

Qwen-Image is Alibaba Qwen's text-to-image model. Hugging Face's safetensors metadata lists 20,430,401,088 BF16 parameters for it, the repo was created on 2025-08-02 and it is licensed Apache 2.0.^[1]

Weights only: file sizes on Hugging Face

Exact sizes from the Qwen/Qwen-Image file listing (1 GB = 1,000,000,000 bytes; split files added together; checked November 4, 2026).^[1]

What this means for your GPU

• At full BF16 precision, the transformer alone (40.9 GB) is bigger than any 24 GB or 32 GB consumer card, so it can't sit fully on one of those GPUs.
• The model card's example loads the pipeline in torch.bfloat16 on CUDA.^[2] Running on smaller cards means CPU offload or a quantized build, which trade speed or quality for memory. See our CUDA out of memory guide for the general fixes.
• Don't have the hardware? You can rent a GPU by VRAM.

Frequently asked questions

How much VRAM does Qwen-Image need?

The weight files total 57.7 GB, and the transformer alone is 40.9 GB in BF16. Qwen does not publish an official VRAM figure, and real use also needs memory for activations.

Can Qwen-Image run on a 24 GB GPU?

Not with every component loaded at full BF16 precision, because the transformer alone is 40.9 GB. It needs CPU offload or a quantized build.

How many parameters does Qwen-Image have?

Hugging Face lists 20,430,401,088 BF16 parameters for the model.

Every number on this page links to its source.