According to IT之家 and DoNews, Alibaba's Qwen team open-sourced BlenderWorld on October 14, a benchmark for evaluating visual agents on real 3D creation tasks in Blender; it is under the MIT license, with code and dataset published on GitHub, Hugging Face and ModelScope. The benchmark contains 223 tasks and 2,651 criteria across 9 major categories and 22 subcategories, ranging from products and architecture to natural environments, characters, animation and scientific visualization, and it requires an editable .blend project file rather than just a rendered image. Evaluation renders submissions from 8 adaptive viewpoints and combines scene checks, task-specific binary criteria and 8 overall visual quality scores, weighting rule compliance and visual quality at 50% each; the project uses Blender 4.2.10 and requires Linux and Python 3.10 or later.