Spectral Labs develops AI CAD benchmark and reports Solidworks exam success

AI-generated summary

Spectral Labs reports that its AI agent has passed the Solidworks Professional exam, autonomously building, modifying and assembling parts and mechanisms in less than three hours. The company claims this makes its system the first AI to achieve an industry-standard CAD certification for mechanical engineers. It says the agent has passed the exam multiple times, although it has not yet achieved a perfect score. Founded in 2022, Spectral Labs describes itself as a spatial intelligence company developing foundation models for engineering design. Its system combines SGS-Mk2, a foundation model for physical engineering, with Quartz, its CAD execution environment. SGS-Mk2 reasons about the geometry, topology and kinematics of parts and assemblies. The company describes the model as lightweight but says it requires orchestration by a vision-language model, or VLM. A single set of model weights is conditioned and accessed through different tools within Quartz, with each tool defined around the requirements of a particular problem type. Alongside the certification result, Spectral Labs evaluated SGS-Mk2 against frontier models using 45 internally prepared CAD problems: 20 inspired by the Solidworks Associate certification, CSWA, and 25 by the Professional certification, CSWP. The CSWP-inspired problems covered all three exam segments, comprising 10 part-modeling tasks, five part-editing tasks and 10 assembly tasks. An answer counted as correct when the relevant volume, center of mass or distance was within 0.1% of the accepted value. Spectral Labs says no exam-specific training was undertaken and reports that using SGS-Mk2 reduced per-problem costs by 46–87% on problems involving geometric reasoning. CEO Rahul Iyer presents the work as a response to a shortage of long-horizon benchmarks for CAD generation, editing and understanding. The Solidworks Professional exam provides a multi-hour, multi-task evaluation rather than a collection of isolated operations. The article also notes that testing exposed issues with the exam itself and exercises that stumped other AI agents, but it does not detail those findings. The performance and cost results are reported by Spectral Labs, and the internally prepared comparison problems should be distinguished from the company’s reported success on the certification exam. Beyond the benchmark, Spectral Labs says it is expanding its work with manufacturers and growing its team. Its stated objective is to automate engineering workflows end to end, spanning design and analysis through manufacturing and sourcing, with a focus on reducing development time and production costs. The exam results demonstrate the CAD capabilities the company reports today; the broader workflow automation remains the direction of its ongoing work, rather than an outcome established by the certification exercise.