The 7B parameter Qwen-Image-2.1 model from Alibaba outperforms all other open source entries in the Image Edit and Text-to-Image Arenas. It scored 1367 points in the Image Edit category, placing it 16th overall and three points behind GPT-Image-1.5-high-fidelity. The architecture enables both image generation and instruction-based editing, featuring native RGBA transparency and support for up to 10 reference images.

Technical integrations include day-0 support for vLLM-Omni, SGLang-Diffusion, and ComfyUI, while local deployment via Unsloth GGUFs allows the model to run on 12GB of VRAM. The model ships with a strict non-commercial license that prohibits use in monetized content. Optimized FP8 versions utilizing RAM offloading can fit in under 8GB of VRAM.