SGS-1: a foundational model for CAD (www.spectrallabs.ai)

Today’s announcement introduces SGS-1, a foundation model that generates fully manufacturable, parametric 3D CAD geometry — producing B-Rep parts exportable as STEP files that can be opened and edited in traditional CAD software. Given an image, sketch, “dumb” 3D mesh (e.g., STL) or a partial assembly plus a text prompt, SGS-1 returns editable parametric geometry suitable for engineering workflows: bracket designs that fit into assemblies, conversion of hand sketches and engineering drawings to CAD, and automated reverse-engineering of meshes into STEP files. Unlike prior generative models that create non-editable meshes, SGS-1’s outputs are accurate, watertight solids that preserve parametric intent and manufacturability. Technically, SGS-1 outperforms state-of-the-art baselines (including an LLM-based CadQuery approach attributed to “GPT-5” and the HoLa latent-diffusion image-to-CAD model) on a 75-image benchmark of medium-to-high complexity parts, evaluated by success ratio using distance metrics over 10 runs per example; SGS-1 produced valid reconstructions for almost all but the hardest objects. Limitations include weaker performance on organic, highly curved shapes, very thin structures, limited 3D resolution, and no single-shot full-assembly generation. The team plans multimodal, larger-context models and reinforcement learning with physics simulation to improve spatial reasoning and physical design capabilities.