Alibaba Qwen open-sources Qwen-Image-2.1-Turbo on October 9, cutting 7B image model from 40 to 8 denoising steps

Alibaba's Qwen team released Qwen-Image-2.1-Turbo open weights on October 9, an accelerated checkpoint of Qwen-Image-2.1 that cuts denoising steps from 40 to 8 for both text-to-image generation and image editing, roughly 5x faster than the base model. The visual generator remains a 7B parameter single-stream DiT paired with a Qwen3-VL 8B text encoder. Output runs up to 2048x2048 square and 2752x1536 at 16:9. The hosted API on Alibaba Cloud PAI-EAS costs CNY 0.10 per image at 120 RPM, 60% cheaper than the CNY 0.25 Pro tier. The model supports up to 10 reference images, native RGBA transparent output, and ships under the Qwen Research License, with weights on ModelScope and Hugging Face; commercial self-hosting requires separate permission. Qwen notes in the model card that the 8-step schedule is baked into the weights, so setting num_inference_steps does not override the default configuration, and callers must pass an explicit sigmas argument. The base Qwen-Image-2.1 scores 60.28 on Qwen-Image-Bench, the highest open-weight score Qwen claims.