Build BenchCAD With Us
BenchCAD is an open, execution-graded effort to measure — and push — how well AI designs real hardware. It grows two ways: every model on the leaderboard maps where the frontier stands, and every new part family raises the bar it has to clear. Contribute either and you help set the standard — every accepted contribution is credited, and major contributors are recognized in the BenchCAD 2.0 paper.
Submit a model to the leaderboard
Run the full benchmark
Run BenchCAD on the public split with your model — see Setup for install and the exact benchcad.py command. Raw predictions are written per task.
Build submission.jsonl
Collect the model's raw outputs as submission.jsonl — one {record_id, prediction} per record, not scores. Sanity-check locally with tools/regrade.py.
Open a Model result submission
Open an issue on the code repo with the Model result submission template and attach submission.jsonl + your run config. We re-grade with the official scorer — voxel IoU for geometry, ratio accuracy for QA — before listing.
Contributors & Hall of Fame
BenchCAD is built by its contributors. The core team below — and a growing wall for everyone who lands a part family, a model submission, or a new benchmark task.
Core team
Community contributors
Land a part family or a model submission and you join the wall, and major contributors are recognized in the BenchCAD 2.0 paper.
Questions, ideas, or want to claim a family before you build it? Open an issue on the 2.0 repo or join the Discord.