CAD-GPT is a multimodal LLM fine-tuned to generate CAD construction sequences from both text and image inputs. Focuses on spatial reasoning — understanding 3D relationships from 2D projections.