A unified architecture allows for the simultaneous generation and editing of images using a single checkpoint that supports native 2K resolution. Alibaba's Qwen-Image-2.1 integrates RGBA layers for transparent image creation and supports instruction-based editing using up to 10 reference images in a single pass. The 7B parameter open-weight model launched with day-0 support in ComfyUI, vLLM-Omni, and SGLang-Diffusion.

Operating on a single NVIDIA RTX 4090 24GB GPU with CPU offload, the system produces 1024x1024 pixel images in 18.7 seconds and modifies images in 21.7 seconds. The model employs a 7.1B DiT paired with a Qwen3-VL-8B vision-language model to improve text rendering and portrait aesthetics. It is released under a strict non-commercial research license and supports OpenAI-compatible APIs.