# BlenderWorld: Benchmarking Visual Agents for 3D Creation from Real-World References (text-only subset) (BlenderWorld (text-only))

> The 78 text-brief tasks of BlenderWorld, a benchmark where agents work inside Blender to turn real-world 3D creation requests into editable Blender projects and requested exports.

Canonical page: https://benchlm.ai/benchmarks/blenderworld-text-only

- Category: [Coding](/coding)
- Last updated: November 13, 2026

## About BlenderWorld (text-only)

- Year: 2026
- Tasks: 78 text-brief 3D creation tasks in Blender
- Format: Task score (0-100), half rubric and half visual quality
- Difficulty: Long-horizon 3D modeling and scene building
- Paper: [BlenderWorld: Benchmarking Visual Agents for 3D Creation from Real-World References](https://arxiv.org/abs/2610.23986)

BlenderWorld (October 2026, Alibaba's Qwen team with Zhejiang University, CUHK, Fudan and NTU) has 223 tasks and 2,651 binary criteria across products, architecture, environments, characters, animation and scientific visualization; 78 tasks give a text brief only and the rest add images or video. Each task score combines rubric satisfaction and quality across eight visual dimensions. Baidu reports ERNIE 5.5 Flash Preview at 79.0 on the text-only subset in its November 2026 launch, with Claude Code 2.1.220, Blender 4.2.10, a 24-hour timeout and Qwen3.8-Max as judge, next to its own runs of Claude Opus 5.5 (86.1), GPT-6.1 Sol (85.3), DeepSeek V4.1 Flash 1017 (79.2), Hy4 (76.8), GLM-5.3-Flash (71.4) and GLM-5.3 (69.7). BenchLM stores these Baidu-run rows as display-only; the official BlenderWorld leaderboard's text column comes from the authors' own runs and is not mixed in.

BlenderWorld (text-only) is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (7 models)

None of the 7 results on this page has a documented run by the benchmark or an independent evaluator: 1 reported by the lab that built the model and 6 reported in another publisher’s document.

| Rank | Model | Creator | Score | Who ran it |
|------|-------|---------|-------|-----------|
| 1 | [Claude Opus 5.5](/models/claude-opus-5-5) | Anthropic | 86.1% | Lab-reported, from another publisher’s document |
| 2 | [GPT-6.1 Sol](/models/gpt-6-1-sol) | OpenAI | 85.3% | Lab-reported, from another publisher’s document |
| 3 | [DeepSeek V4.1 Flash 1017](/models/deepseek-v4-1-flash) | DeepSeek | 79.2% | Lab-reported, from another publisher’s document |
| 4 | [ERNIE 5.5 Flash Preview](/models/ernie-5-5-flash-preview) | Baidu | 79.0% | Reported by Baidu |
| 5 | [Hy4](/models/hy4) | Tencent | 76.8% | Lab-reported, from another publisher’s document |
| 6 | [GLM-5.3-Flash](/models/glm-5-3-flash) | Z.AI | 71.4% | Lab-reported, from another publisher’s document |
| 7 | [GLM-5.3](/models/glm-5-3) | Z.AI | 69.7% | Lab-reported, from another publisher’s document |

## FAQ

### What does BlenderWorld (text-only) measure?

The 78 text-brief tasks of BlenderWorld, a benchmark where agents work inside Blender to turn real-world 3D creation requests into editable Blender projects and requested exports.

### Which model scores highest on BlenderWorld (text-only)?

Claude Opus 5.5 by Anthropic currently leads with a score of 86.1% on BlenderWorld (text-only).

### How many models are evaluated on BlenderWorld (text-only)?

7 AI models have published results on BlenderWorld (text-only) in the BenchLM catalog.

### Does BlenderWorld (text-only) affect BenchLM's overall score?

Not directly. BlenderWorld (text-only) is still displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Compare Top Models on BlenderWorld (text-only)

- [Claude Opus 5.5 vs GPT-6.1 Sol](/compare/claude-opus-5-5-vs-gpt-6-1-sol)
- [GPT-6.1 Sol vs DeepSeek V4.1 Flash 1017](/compare/deepseek-v4-1-flash-vs-gpt-6-1-sol)
- [DeepSeek V4.1 Flash 1017 vs ERNIE 5.5 Flash Preview](/compare/deepseek-v4-1-flash-vs-ernie-5-5-flash-preview)
- [ERNIE 5.5 Flash Preview vs Hy4](/compare/ernie-5-5-flash-preview-vs-hy4)