We Tried Text-to-CAD Tools So You Don't Have To — Here's Where They Actually Break

https://ezremove.ai/ai-video-generator/The promise sounds almost too convenient: describe a part in plain English and receive a production-ready STEP or STL file a minute later. The combination of LLMs, CadQuery, OpenSCAD, and cloud-based CAD platforms such as Zoo.dev has made text-to-CAD accessible not only to research labs, but also to everyday engineers. Zoo, for example, can generate editable parametric models from text prompts.

We tested several AI CAD generation tools on typical mechanical parts, ranging from simple brackets to objects with multiple bodies, holes, fillets, and interdependent dimensions. The results were more interesting than the usual debate about whether AI will replace engineers. The technology is already useful — sometimes impressively so — but it reaches its limits much sooner than product demos suggest.

What “Text-to-CAD” Actually Promises

In its most common form, a text-to-CAD system receives a written prompt, converts it into CadQuery or OpenSCAD code, executes that code, and exports the resulting geometry as a STEP, STL, or another CAD-compatible format. More closed platforms may generate models directly, but the underlying idea remains the same: the user describes the geometry, and the model creates a sequence of operations instead of requiring manual work in a CAD interface.

This is not hype without substance. In one study, researchers expanded a dataset with approximately 170,000 CadQuery annotations and fine-tuned six open-source LLMs. The best-performing model achieved a 69.3% top-1 exact-match accuracy, compared with a baseline result of 58.8%.

There is, however, an important qualification. That result was not achieved because any general-purpose chatbot can suddenly design mechanical parts. It depended on LLM fine-tuning using specialized CAD data. In other words, CadQuery AI and OpenSCAD AI systems perform best when the model understands a specific construction language, common modeling operations, and the structure of parametric geometry.

For companies that need a model trained on their own parts library, naming conventions, and internal engineering standards, this becomes a large language model development problem rather than a matter of finding a better prompt. The quality of AI generated CAD models depends on training data and model customization at least as much as it depends on model size.

Where It Actually Works

We obtained the best results with parts that could be expressed as a short sequence of predictable operations: create a sketch, extrude a profile, add holes, and repeat a feature using a pattern. Bolts, spacers, mounting plates, simple enclosures, and L-shaped brackets are almost ideal for this approach.

In one test, we asked a system to create a mounting plate with four holes, a central cutout, and configurable values for width, thickness, and hole spacing. The first version could already be exported, and after one follow-up prompt, the model converted the main dimensions into variables. This is more than a toy demonstration. For a family of similar parts, a script like this could save a meaningful amount of engineering time.

Rapid prototypes also work well when obtaining the overall shape matters more than producing flawless manufacturing geometry. For a visual mockup of a sensor enclosure or an early assembly clearance check, it may be enough to have an LLM generate a CAD model with approximately the right dimensions and let an engineer correct the details afterward.

This is why engineering feedback around code-based CAD is often positive. Parametric automation is especially useful for catalogs of similar components where dimensions are the main variables. One CadQuery user, for example, described automatically generating approximately a thousand electronic components that differed primarily in their parameters. In these scenarios, AI is not inventing geometry. It is accelerating the implementation of logic that is already understood.