Alibaba Qwen open-sources BlenderWorld for AI 3D modeling assessments
October 19, 2026 · 2 min read
On October 14, Alibaba Qwen team open-sourced BlenderWorld, a benchmark designed for evaluating visual agents performing real 3D creation tasks in Blender. The project is licensed under MIT, with code and datasets published on GitHub, Hugging Face, and ModelScope.
Unlike typical "generate an image" tasks, BlenderWorld requires AI to deliver an editable .blend project file and produce rendered images, videos, or model export files as needed. In this evaluation, models must not only create visually appealing scenes but also build 3D environments that designers can modify, including materials, geometry, lighting, and camera setups.
The benchmark features 223 tasks and 2,651 standards, covering nine categories and 22 subcategories such as products, architecture, natural environments, characters, animations, and scientific visualizations. Among these, 78 tasks provide only text descriptions, 130 tasks include images and text, and 15 tasks come with videos and text; additionally, 59 tasks have specifications like dimensions, materials, or camera parameters.
The evaluation method is stringent: the system renders the submitted Blender projects from eight adaptive perspectives, using scene checks, a binary scoring system based on the 2,651 task-specific criteria, and eight overall visual quality ratings. The final score is weighted equally between rule compliance and visual quality.
The project currently provides a complete installation, generation, packaging, and evaluation process, allowing developers to integrate any generation framework into the tests. The repository indicates it uses Blender version 4.2.10 and requires a Linux environment with Python 3.10 or higher; the evaluation phase also necessitates configuring a model service with visual capabilities.