AI-generated CAD just got a harder problem to solve: building assemblies, not just parts.
Researchers have published AssemCAD, a framework that takes a natural language description and produces a mechanically valid, multi-component CAD assembly. Rather than spitting out a single block of CAD code, AssemCAD first builds what the team calls an "Assembly Specification" — a structured blueprint that defines typed parts, connection ports, assembly relations, and engineering axioms. A separate library then executes those relations through deterministic geometric transformations and checks them against actual geometric evidence, a process called B-Rep validation. The system can also generate reusable parametric component factories on demand, meaning the same part template can be resized and reused across different assemblies.
The distinction from existing Text-to-CAD work matters because single-part generation is essentially a solved problem for simple geometries, while assemblies require reasoning about how parts fit, move, and constrain one another. Getting that wrong produces models that look plausible on screen but would fail the moment someone tried to manufacture or simulate them. AssemCAD's axiom-grounded approach makes each assembly relation traceable to an engineering principle, which means the output can be inspected and corrected rather than treated as a black box.
On the AssemBench benchmark, AssemCAD outperformed code-centric baselines on both assembly preservation and physical validity, and it generalizes across different underlying language models. The gap between "generates CAD" and "generates manufacturable CAD" has narrowed, but production pipelines will want to see this tested on the messier geometries that real mechanical engineering actually involves.