Qwen-Image-2.1-Turbo open-sourced, image generation steps cut from 40 to 8.

Dongcha Beating AI News Flash: Alibaba’s Qwen has launched Qwen-Image-2.1-Turbo, an accelerated version of Qwen-Image-2.1 that retains the original’s 7-billion-parameter image generation architecture. The official original model uses 40 denoising steps, while the Turbo version cuts this to 8 steps – an 80% reduction. Denoising refers to the process where a model gradually generates an image from random noise. However, an 80% step reduction does not translate to an equivalent 80% time cut. The new model still produces 2K images, supporting text-guided editing, multi-image reference, and transparent background output. Developers can run it locally via Hugging Face’s Diffusers library, which automatically applies the recommended 8-step sampling setting. APIs for both Turbo and Pro versions are now live. Based on Alibaba Cloud’s Beijing Bailian region original pricing, Turbo costs 0.1 yuan per image, while Pro is priced at 0.25 yuan – making Turbo 60% cheaper. Some users have reported significantly faster generation speeds, but quality reviews are mixed. One user testing on an RTX 5090 found Turbo over-sharpens skin textures in generated characters. There is currently a lack of sufficient independent evaluations to confirm its image quality matches the original. Notably, Turbo is licensed under the Qwen Research License: while model weights are downloadable, free use is limited to non-commercial research and evaluation. Commercial development using the weights requires a separate authorization application.