Practical 3D Printing and Automated Rendering Workflow Built with Qwen 27B and CadQuery
A developer created a local multi-modal AI workflow combining a Qwen 27B model with CadQuery, Blender, and ComfyUI to generate 3D-printable objects and advertisement renders. The system successfully designed functional models for a self-watering plant pot and a MagSafe iPhone stand, providing multi-view technical sheets alongside rendered product scenes. This project demonstrates how open-weight LLMs can automate technical hardware design by converting natural language prompts into parametric CAD code instead of just generating standard text or images. It highlights a practical application of agentic tool-use that bridges the gap between conversational AI, parametric modeling, and local digital manufacturing. The hardware setup runs a quantized Qwen 27B model in NVFP4 format on an RTX 5090 GPU, paired with Qwen Image 2.1 in ComfyUI on an RTX 3080 GPU. Using custom skills integrated into an agent harness, the LLM executes CadQuery scripts to generate physical 3D geometries and automatically calculates estimated filament usage for printing.
## BACKGROUND
CadQuery is an open-source Python library designed to build parametric 3D CAD models via code, enabling models to be updated dynamically without using a traditional graphical editor. NVFP4 is a low-precision 4-bit floating-point format developed for NVIDIA GPUs to reduce memory footprint, often used alongside Quantization-Aware Training (QAT) to preserve model performance.