Alibaba’s Qwen team released Qwen-Image-2.1, a 7B open-weight image generation and editing model with native RGBA transparency, up to 10 references, and a research license that bars commercial use without a separate grant. The unified text-to-image generation and editing model’s visual generation stack uses about 7 billion parameters across 32 single-stream DiT layers, targeting 2K-class outputs, refined typography and portrait lighting, and strong quality-per-compute on consumer GPUs such as an RTX 3090-class card. Weights and demos are posted on Hugging Face, ModelScope and GitHub. Native RGBA support lets the same checkpoint generate or edit transparent layers and pull subjects from photographs without a separate matting model, while editing accepts up to ten reference images for identity-preserving group composites, virtual try-on and room restyling, with circles, painted marks or external masks localizing changes. Mixed-granularity attention and prefix KV-cache reuse aim to keep multi-reference edits interactive. Supported aspect ratios stretch from square 2048 to ultrawide and tall phone frames. On Qwen’s own leaderboard the team says the 7B generator outranks most closed image systems it compared, though independent public benches are still sparse. Diffusers pipelines cover text-to-image, edit and RGBA sticker-style prompts, with CPU offload hooks and seedable generators. Qwen-Image-2.