LLMForge AI framework generates CAD designs from text with 99% success

Researchers have introduced LLMForge, a multi-model text-to-CAD framework that leverages large language models (LLMs) and vision-language models (VLMs) to automatically generate parametric 3D mechanical designs from natural-language specifications. The framework integrates JSON-schema validation, analytic feature scoring, mesh synthesis, and multi-round iterative refinement under two critique regimes: IterTracer uses a Phong-shaded ray-trace renderer with geometric metrics (silhouette IoU, hole visibility, edge clearance, aspect-ratio conformance) for lightweight feedback; IterVision replaces analytic scoring with a Qwen2.5-VL-72B VLM that evaluates rendered views via chain-of-thought reasoning.

The study benchmarks seven foundation models (DeepSeek-V3.2, Qwen3-235B-A22B, Llama-3.3-70B, Gemma-3-27B, GLM-4.5, MiniMax-M2.1, and INTELLECT) on 97 engineering problems spanning plates with holes, multi-feature boxes, flanged cylinders, and L-brackets. Under IterTracer, the top four models form a tight cluster with overall mean scores of 0.885–0.890 and 98.97% mesh success, showing compact instruction-tuned models can match much larger systems. IterVision achieves 100% watertight mesh generation on the leading model, though rotationally symmetric geometries like cylinders expose divergence between visual and semantic scoring. The work highlights implications for scalable automated mechanical design in industrial workflows.