CADmium fine-tunes code-LLMs (1.5B–14B) for text-to-CAD sequential design, generating editable parametric CAD construction sequences from natural-language descriptions. It ships with a 170k-model annotated dataset pairing CAD programs with rich captions. A first-in-series domain foundation model from Mila that treats CAD as a program synthesis problem. (Derivative — the models are fine-tuned from existing code LLMs, so no from-scratch parameter count is claimed here.)

Paper

Dataset

Size 170k models
sciencecodedatasetopen-weight