QwenLM releases BlenderWorld benchmark for 3D creation agents
QwenLM released BlenderWorld, an open-source benchmark with 223 tasks across 9 categories for evaluating visual agents on 3D creation in Blender. The dataset includes 2,651 binary criteria and supports text, image, and video references. The source release was published on 2026-10-14.
China context
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can use the public dataset and evaluation protocol to benchmark their own 3D creation agents against the same tasks and criteria, enabling direct comparison with Qwen's reported performance.
- For investors
- The release of a standardized benchmark for 3D creation agents signals Qwen's investment in the 3D generation space, which may attract developer mindshare and ecosystem growth around its tools.
BlenderWorld is a benchmark for visual agents that create editable 3D scenes from real-world references. It contains 223 tasks across 9 categories and 22 subcategories, with 2,651 binary criteria. Tasks include text, image, and video references, and 59 tasks include specification attachments. The evaluation pipeline uses 8 adaptive views, scene inspection, and a weighted binary rubric combined with 8 visual-quality ratings. The benchmark is designed for Blender 4.2.10 and supports any generation harness, with optional Harbor integration.
The benchmark evaluates agents on producing editable Blender projects from references, using a two-part scoring system: binary criteria (weighted 1-3) and eight visual-quality dimensions, combined 50:50. The public evaluator supports CPU Cycles and accepts submissions from any generation harness, while the paper setup uses GPU Cycles and Blender MCP. The renderer selects eight adaptive views at 1024×1024 with 64 samples, allowing up to two corrective re-renders per view.
Developers of 3D generation agents outside China gain a standardized, reproducible evaluation suite for Blender-based creation tasks, reducing the cost of benchmarking and enabling direct comparison with Qwen's reported results. The benchmark's open dataset and evaluation protocol lower the barrier for teams to test their models on realistic 3D creation requests without building custom infrastructure.
For AI labs and tooling companies targeting 3D content creation, BlenderWorld provides a concrete benchmark to validate agent capabilities and differentiate products. The open-source nature allows integration into internal evaluation pipelines, potentially accelerating development cycles for 3D generation agents.
A verifiable next signal is whether independent teams submit results to the BlenderWorld leaderboard or publish comparisons using the public evaluator, which would indicate adoption beyond Qwen. Another signal is the release of agent checkpoints or training data derived from the benchmark tasks.