Alibaba's Qwen team ships 7B image model that works in 8 steps

First reported by RuntimeWire at · Updated · 2 sources

Qwen-Image-2.1-Turbo is an accelerated version of the open-weight Qwen-Image-2.1 and both generates and edits images. It needs 8 denoising steps, compared with 40 by default for the base model, on the same 7B architecture. According to RuntimeWire, commercial use requires a separate license under Qwen's research license.

  • Weights are downloadable, and a hosted API option is also available, according to MarkTechPost.
  • The model is a 7B checkpoint covering image generation and editing.
  • Step count falls fivefold compared with the base model.

Covered by 2 publishers within 2 hours of the first report.

Reporting2