MUSE (Text-to-CAD) - Geometric Validity: leaderboard
Geometric validity in percent: the share of 106 multi-part product design specifications that models turn into CadQuery code (MUSE text-to-CAD benchmark, 2026), evaluated as a funnel in which a failed stage scores zero downstream whose exported model is watertight, manifold, free of self-intersections and free of overlapping parts. Higher is better.
Metric: Geometric validity (%): share of the 106 design specifications whose exported STEP model passes all four binary geometry checks (watertight, manifold, self-intersection free, overlap free), one CadQuery script per design specification, executed in a sandbox; funnel evaluation, so a sample that fails a stage scores zero on every later stage; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 68.87 | #54 |
| 2 | Gemini 3.1 Pro (Preview) | 58.49 | #99 |
| 3 | Claude Opus 4.7 | 58.49 | #85 |
| 4 | GLM-5.1 | 27.36 | #198 |
| 5 | Claude 3.7 Sonnet | 23.58 | #470 |
| 6 | Llama 3.1 70B | 23.58 | #1499 |
| 7 | Qwen 2.5 72B | 20.75 | #1188 |
| 8 | GPT-4o | 13.21 | #751 |
| 9 | Qwen 3.5 122B A10B | 13.21 | #346 |
| 10 | MiniMax-M2.7 | 10.38 | #539 |
| 11 | MiniMax-M2.5 | 10.38 | #600 |
| 12 | MUSE Qwen3.6-Coder (checkpoint unspecified) | 10.38 | |
| 13 | Qwen 3.6 35B A3B | 5.66 | #481 |
| 14 | GLM-4.7 Flash | 5.66 | #1349 |
| 15 | Llama 3.1 8B | 1.89 | #5196 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=muse-text-to-cad-geometric-validity · How It Works · Data refreshed daily, snapshot 2026-11-21.