CADEngBench - Assembly End-to-End (A2): leaderboard
The end-to-end assembly task of CADEngBench (MBZUAI, 2026): the rank-1 joint must be correct and its predicted joint frame, placement and allowed and blocked motions must pass scripted kinematic tests. Percentage of 120 tasks passed; higher is better.
Metric: End-to-end pass rate (%; share of 120 assembly relation tasks between two source B-Rep bodies; one response per task whose rank-1 joint is correct and whose predicted joint frame, placement and allowed and blocked motions pass the scripted kinematic tests). Source: arxiv.org. Saturation forecast: Around 2029. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) (Minimal) | 15.8 | #144 (Gemini 3 Flash (Preview)) |
| 2 | Claude Sonnet 4.5 | 12.5 | #250 |
| 3 | GPT-5.2 (2025-12-11) (Non-reasoning) | 11.7 | |
| 4 | Llama 4 Maverick | 10.8 | #1099 |
| 5 | Qwen 3.5 35B A3B (Non-reasoning) | 10.8 | #482 (Qwen 3.5 35B A3B) |
| 6 | GLM-4.6V (Non-reasoning) | 8.3 | #701 (GLM-4.6V) |
| 7 | Kimi K2.5 (Non-reasoning) | 7.5 | #256 (Kimi K2.5) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cadengbench-assembly-end-to-end-a2 · How It Works · Data refreshed daily, snapshot 2026-11-13.