Qwen published the Qwen-Image-2.1 weights on 20 September under the Qwen Research License, which covers research and evaluation rather than production, and opened the Pro and Turbo APIs on Alibaba Cloud Model Studio on 9 October. The architecture is a 32-layer single-stream diffusion transformer with 7B parameters, paired with a Qwen3-VL 8B text encoder and a 64-channel RGBA variational autoencoder. Text and reference images are encoded once at the first denoising step and reused for every later step, which the team describes as the basis for handling multi-image inputs quickly. The model accepts up to 10 reference images, performs region edits from a mask or annotation, renders typography, and outputs a native alpha channel, so a cutout arrives as an RGBA file rather than requiring a separate matting pass. Default output is 2048 by 2048 with presets from 9:16 to 16:9. The Turbo checkpoint keeps the same 7B architecture at 8 denoising steps instead of 40. Pricing in Singapore is 0.016 US dollars per image at 120 requests per minute for Turbo, and 0.04 dollars per image at 20 requests per minute for Pro.
Our takeThe alpha channel is an interface change rather than a quality claim: removing a matting step changes how many tools a pipeline needs, and that is measurable. Caching the first-step encodings is what makes ten references affordable, since extra images stop costing extra denoising passes. Two things to test before trusting it: whether edits outside the masked region stay untouched at 2K, and whether the Turbo tier holds composition on multi-reference inputs, where step reduction usually shows first.