Home/AI NOW/NVIDIA releases open Qwen-Image-Flash model that generates in four steps

ImageOPEN SOURCE

NVIDIA releases open Qwen-Image-Flash model that generates in four steps

NVIDIA has published Qwen-Image-Flash on Hugging Face, a DMD2-distilled Qwen-Image checkpoint designed to generate images in four denoising steps.

07/23/20261 sources reviewed

Quick summary

It retains the base Qwen-Image architecture and parameter count while using four denoising steps.
It was validated with English prompts at 1024×1024 and supports Diffusers, SGLang Diffusion, vLLM-Omni and TensorRT-LLM.
Commercial and non-commercial use is supported, but checkpoint memory requirements are unchanged and no safety checker is included.

WHAT HAPPENED

What happened?

NVIDIA has published Qwen-Image-Flash on Hugging Face, a DMD2-distilled Qwen-Image checkpoint designed to generate images in four denoising steps.

Fewer steps can improve generation speed, but this is not a smaller model. Deployers should separately test GPU memory, non-English prompt quality and their own content safeguards.

WHY IT MATTERS

Why does it matter?

Reducing the number of generation steps can cut latency for live previews and high-volume creative workflows.

WHO SHOULD CARE

Who should care?

Image-generation developersCreatorsOpen-model researchers

AIZIGOO VIEW

AIZIGOO view

Capabilities and limitations follow NVIDIA's model card released on July 23, 2026.