gpt-6IMAGE

Qwen open-sources BlenderWorld: GPT-6 Astra tops at 79.9 but only 42.2% of tasks fully pass

A new 3D creation agent benchmark of 223 tasks requiring editable .blend files; GPT-6 Astra leads at 79.9 overall but spatial consistency and physics remain weak spots.

SuSu_酥酥👅

@NFT_ChenArchived 4 days agoOriginally posted 2026.10.24

Qwen open-sources BlenderWorld: GPT-6 Astra tops at 79.9 but only 42.2% of tasks fully pass (image 1 of 2)

Qwen open-sources BlenderWorld: GPT-6 Astra tops at 79.9 but only 42.2% of tasks fully pass (image 2 of 2)

SuSu_酥酥👅 @NFT_ChenOriginally posted 2026.10.24

Open original on X

🔥2D好看≠3D成立！Qwen已开源BlenderWorld，首个真实3D创作Agent评测基准来了！ 223个生产级任务，要求Agent根据文本/图片/视频/规格文件，在Blender里交付可编辑.blend工程，而非一张漂亮渲染图。实测最强GPT-6-Astra综合79.9分，但仅42.2%任务通过全部验收，空间一致性、物理关系、细节遵循仍是硬伤。核心亮点： 🔹2651条二值验收标准 + 多视角渲染 + 场景结构检查双轨评分 🔹必须交付原生可编辑工程（没.blend直接0分） 🔹支持任意Agent/工具链接入，Harbor可选适配快速上手： 1️⃣git clone https://github.com/QwenLM/BlenderWorld && cd BlenderWorld && pip install -e . 2️⃣bw download --output data/BlenderWorld 拉取数据集 3️⃣bw prepare 准备任务 → 你的Agent生成.blend 4️⃣bw pack-submission 打包提交 5️⃣bw evaluate 一键跑评测 #Qwen #BlenderWorld #3DAI #OpenSource #Agent #Alibaba

This post quotes another post: x.com/qwendevs/status/2113474673606631844

x.com/nft_chen/status/2113861423411667245Archived 2026.10.25 10:04