动察 Beating AI News Flash: Alibaba's Tongyi Qianwen releases Qwen-Image-2.1-Turbo and opens model weights. This is an accelerated version of Qwen-Image-2.1, retaining the 7-billion-parameter image generation architecture. The original official examples use 40-step denoising, while Turbo reduces it to 8 steps, a 80% reduction in steps. Denoising is the process by which the model progressively generates an image from noise. However, an 80% reduction in steps does not mean actual time consumption is also reduced by 80%.The new model can still generate 2K images, support text-instruction image editing, multi-image reference, and transparent background output. Developers can run it locally via Hugging Face's Diffusers library, automatically using the recommended 8-step sampling settings. The Turbo and Pro APIs are also now live. At Alibaba Cloud Bailian Beijing region list prices, Turbo charges ¥0.1 per image, while the Pro version is ¥0.25, making the former 60% cheaper.Some users have reported noticeably faster generation speeds, but image quality reviews are mixed. A user testing on an RTX 5090 found that Turbo over-sharpens skin textures in generated human figures. There is currently a lack of sufficient independent evaluations to prove its image quality is on par with the original version.It should be noted that Turbo uses the Qwen Research License. Although the model weights can be downloaded, free use is limited to non-commercial research and evaluation. Commercial development using the weights still requires a separate authorization application.