Open-Source AI System Turns Text Descriptions Into 3D Models for Under $2

A research team at Tsinghua University has released an open-source system called MAC (Multi-Agent CAD) that turns text descriptions into complete, printable 3D models. The key claim is efficiency: the system reportedly uses 99% fewer computational tokens than existing text-to-CAD tools, cutting the cost of a ten-part design run from about 16 euros down to around 1.25 euros.

How It Works

MAC splits the work across four specialized agents connected by a LangGraph state machine. A Spec Planner converts the prompt into a structured design specification. A Geometric Architect turns that into a geometric plan with sketches and build steps. A Python Coder then writes the actual CAD code for the build123d library without needing another language model call. A validation step checks the STEP geometry and STL mesh, fixing errors automatically if needed.

The efficiency gain comes from what the agents exchange. Earlier systems feed the entire conversation history back into the language model at every step, which drives up token use fast. MAC agents only pass structured intermediate results to each other. For ten test tasks with 141 total geometric features, the comparison system used 103.9 million tokens and 1,307 API calls. MAC used 896,340 tokens and 50 calls. The success rate also ticked up, from 97.9% to 99.3%.

What It Can Actually Print

The system handles multipart assemblies that come out of the printer ready to assemble, with deliberate 0.4 to 1 millimeter gaps between parts. In one test, a spiral staircase model, the system recognized on its own that the specified geometry would have broken during printing because parts would touch at only one point. It adjusted the overlap automatically.

MAC is released under the MIT license on GitHub. It runs from the command line or through a browser interface with a 3D preview, and it accepts various language models via an OpenAI-compatible interface. The team lists dental frameworks, aerospace parts like nozzle guide vanes and blisks, and consumer wearable designs as example use cases.