OmniCAD - Position Accuracy: leaderboard
OmniCAD (Pennsylvania State University and Shanghai Jiao Tong University, 2026) evaluates vision-language models on assembly-aware 3D spatial reasoning, reconstructing component poses of robotics CAD assemblies from color-by-component views. Position accuracy is the share of components whose predicted translation is within 10mm of the ground truth, in percent, higher is better.
Metric: Pos@10mm (%; share of components whose predicted translation is within 10mm of the ground truth; over OmniCAD component-level CAD assembly generation; color-by-component views input (ColorView)). Source: arxiv.org. Saturation forecast: Not forecast. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 4 Maverick | 22.2 | #1105 |
| 2 | Qwen 3.5 9B | 20.2 | #803 |
| 3 | Claude Opus 4.8 | 19.5 | #56 |
| 4 | Ministral 3 14B | 16.1 | #1852 |
| 5 | GPT-5.5 | 15.7 | #48 |
| 6 | Qwen 3.7 Plus | 15 | #203 |
| 7 | GPT-5 Mini | 11.4 | #349 |
| 8 | Gemini 3.1 Pro (Preview) | 8.5 | #91 |
| 9 | Gemini 2.5 Flash Lite | 8.1 | #958 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=omnicad-position-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.