Alibaba's Qwen team has released Qwen-Image-2.1, an open-weight model that handles both text-to-image generation and instruction-based editing in a single network. The generator runs at around 7 billion parameters, well below the roughly 20 billion of the original Qwen-Image, and ships as a 14.2GB BF16 build or a 7.26GB INT8 version.
The model produces native 2K output, keeps text legible inside generated images, and supports an alpha channel for transparent backgrounds. For editing, it accepts up to ten reference images at once, which the Qwen team says improves fidelity when working with portraits and products rather than relying on a single input photo.
Weights are available without gating on Hugging Face and ModelScope, and the code is on GitHub. ComfyUI added support from launch, and SGLang reported that a single RTX 4090 with 24GB of memory generated a 1024×1024 image in 18.7 seconds and completed an edit in 21.7 seconds, using CPU offload at native precision.
Licensing is more restrictive than the open weights might suggest: several listings describe the terms as a research licence that limits commercial use, so companies planning to deploy the model in a paid product should check the Hugging Face model card before doing so.
Qwen-Image-2.1 continues Alibaba's push to make capable generative models small enough to run on consumer hardware, following similar compression work across its Qwen3 language and Qwen-Image lines this year.
Source: MarkTechPost
