AI news ·
3DCodeVerse releases 2.03 million executable 3D code pairs across six languages, achieves 90% human-preference win rate over Claude design
A new open-source dataset delivers 2.03 million executable 3D code pairs and a verification harness that achieves a 90% human-preference win rate against Claude Opus 5.5 Design.
A new open-source dataset released October 16, 2026, delivers 2.03 million executable 3D code pairs spanning six programming languages. The release includes a multi-agent verification harness that uses visual language models to judge and refine generated code, achieving a 90% human-preference win rate against Claude Opus 5.5 Design.
The project, called 3DCodeVerse, targets a persistent gap in code generation: most models can produce syntactically correct code, but 3D programming requires spatial reasoning that text-only evaluation misses. Each pair in the dataset links a natural language prompt to working code in Blender Python, CadQuery, Three.js, GLSL, OpenSCAD, or URDF. The code has been executed and visually verified, not just parsed.
How the verification harness works
The multi-agent system generates candidate code, executes it, renders the output, and feeds the visual result to a vision-language model for quality assessment. If the render does not match the prompt's intent, the system iterates. This closed-loop refinement adds 15.5 to 29.1 percentage points over single-shot baselines.
Fine-tuning Qwen 3.5-9B on the dataset alone produces stark gains. On 3DCodeBench, performance jumps from 0% to 84.9%. On CadQuery, it climbs from 10.5% to 98.5%. The harness then pushes results higher, reaching the 90% win rate in head-to-head human evaluations against Claude's design-focused model.
Six languages, one verification pipeline
The dataset's language coverage reflects distinct 3D workflows. Blender Python and Three.js serve artists and web developers. CadQuery and OpenSCAD target parametric CAD for engineering. GLSL covers shader programming, and URDF supports robot description formats used in simulation. All six share the same execution-and-render verification loop.
Hugging Face originally reported the release. The full dataset and harness code are available under an open-source license, letting teams fine-tune their own models or integrate the verification step into existing pipelines. For developers building AI code generation tools, the render-based evaluation approach offers a template that extends beyond 3D domains.
Why this matters for creatives, educators, and developers
For anyone who generates 3D assets through code, the bottleneck has always been trust: you cannot ship a script that looks right but produces broken geometry. A dataset of verified, executable pairs removes that guesswork. Fine-tuned models can handle routine CAD scripting or shader authoring, cutting hours of manual debugging. Educators teaching AI-assisted coding now have a concrete example of how execution feedback transforms model reliability - from zero to production-usable in a single fine-tuning run.