2026 JoyIndustrial VisCAD Industrial CAD Intelligence

AI-assisted computer-aided design (CAD) for industrial products involves two challenging phases: part-level and assembly-level generation. Existing specialized CAD models often struggle with generalization due to their narrow training domains, while general-purpose models tend to perform inconsistently. This paper presents VisCAD, a foundation model suite designed to enhance both generalization and CAD capabilities for realistic industrial products. The core of VisCAD is VisCAD-M1, a 27B model that utilizes mid-training and post-training approaches for part-level design generation. Evaluation on the PubCADBench and RealCADBench shows that VisCAD-M1 achieves the highest average part-level score of 0.5540 among the evaluated models, surpassing the strongest frontier model which scored 0.5496. Furthermore, applying VisCAD-M1 as a test-time verifier improves this score to 0.5797, representing a 5 percent relative increase over the previous state of the art. VisCAD also incorporates a domain-specific harness which successfully leverages frontier models for complex assembly generation, demonstrating both quantitative and qualitative advantages over general-purpose solutions. This work illustrates the potential of advanced CAD intelligence for industrial applications, contributing to improved design accuracy and efficiency in a rapidly evolving technological landscape.