VisCAD foundation model suite targets industrial CAD generation across modalities

VisCAD's actual novelty isn't just multimodal input handling (that's table stakes now) but the explicit modeling of assembly-level constraints like mating relations and spatial positioning. Most foundation models treat these as downstream inference problems; VisCAD bakes them into the architecture itself.

This connects directly to the Visual Insensitivity Gap research from early September, which found that vision-language models often ignore their visual input entirely. VisCAD sidesteps that problem by training on domain-specific visual reasoning (CAD geometry, spatial relations) rather than relying on general-purpose VLMs to infer what matters. It also echoes the document VLM work from the same period, which showed that specialized fine-tuning on production data outperforms larger generalist baselines. For CAD, the constraint is even tighter: a model that hallucinates part geometry or violates assembly rules is useless, not just suboptimal.

If VisCAD's part-generation outputs maintain consistency when assembled into multi-part models (zero constraint violations on held-out assembly sequences), that validates the claim that vertical models handle constraint reasoning better than horizontal ones. If the same benchmark fails on novel assembly topologies not seen in training, the model is memorizing rather than reasoning.