Alibaba releases Qwen-Image-2.1 with research-only license

Alibaba’s Qwen team released Qwen-Image-2.1 on Hugging Face and ModelScope, adding a 7B, 32-layer single-stream DiT architecture paired with a Qwen3-VL 8B text encoder. The model uses a 64-channel RGBA VAE and is designed to generate native 2048×2048 images at 40 steps.

The release includes two 9B PE-T2I/PE-I2I prompt-rewriter checkpoints and supports up to 10 reference images, along with mask/circle-based local editing. Reported strengths include panoramas, infographics, virtual try-ons, realistic textures, and typography.

The licensing shift is a major change for users of the earlier Qwen-Image line. Alibaba has moved from Apache 2.0 to a non-commercial Qwen Research License Agreement, meaning commercial users now need a separate agreement.