Trained two 16M-param models to do generative CAD with real physics

debreuil · x · 2026-08-24

Most "AI CAD" demos are the same thing: a frontier LLM writing code translated by a geometry kernel into a 3D B-rep model. connorkapoor argues this path is lossy and lacks the spatial understanding humans apply when designing in CAD.

Instead, building on generative 3D research, he trained two tiny 16M-parameter models on a hyper-specific dataset, introducing FieldGeo: models that encode geometry, physics (FEA, stress) and DFM into the same latent space, reproducing geometries that meet stated cases with the help of a design agent. Today it only handles L-shaped brackets.

Original post →

More from Multimodal

Multimodal channel →