Qwen-Image-2.1-Turbo: High-Speed Image Generation in 8 Steps

What happened

On October 9, the Qwen team released Qwen-Image-2.1-Turbo on Hugging Face — an official accelerated version of Qwen-Image-2.1. While baseline models typically need dozens of denoising steps, the Turbo checkpoint compresses that to 8 steps, with the sampling schedule saved directly into the weights. No manual scheduler tuning is needed; it loads with a single QwenImage21Pipeline call. It defaults to CFG=1 and reuses text and reference-image context across steps via prefix KV caching, pushing latency down further.

Key facts

1. 8-step generation:official accelerated checkpoint with the sampling schedule baked in; CFG=1 by default, with prefix KV caching reusing text and reference-image context.
2. Same 7B architecture:keeps the 7B visual-generation architecture of Qwen-Image-2.1; loads in one line via Diffusers as a drop-in replacement for the baseline weights.
3. Generation + editing:supports image editing up to 2048×2048 plus preset aspect ratios including 16:9 (2752×1536) and portrait posters (1680×2512).
4. License:released under the Qwen Research License, not Apache 2.0 — open weights, separate terms for commercial use.