Which AI designs the best parts?

nurb works with the AI subscription you already have. We give each model the same real part-design jobs and grade the parts by machine: the actual geometry, checked against what was asked and against print physics. No cherry-picking, no vibes. Start from what you subscribe to.

The short answer

No rows on file yet

be the first

the one-liner below runs the benchmark on your own subscription and stages a submission

The tradeoff

Every benchmarked model and effort level, placed by how often its parts print right the first time against how long it takes per part. Up and to the left is better.

The chart appears with the first submitted rows.

The jobs

Follow the specA cable clip with every dimension stated. Can it build exactly what you asked?

Design from a problem“Hold this cable bundle on the wall with one screw.” No shape given: it has to design one that works and prints.

Handle a missing measurementOne dimension nobody measured. Does it guess silently, or handle the unknown the honest way?

The receipts

Per-model detail: the verdict, and every job's score with a tick per attempt. Click a model to open it.

Nothing yet.

Add your model to this page. One line, your own subscription, a wizard for the rest. A single run counts; it pools with everyone else's. curl -fsSL https://nurb.dev/bench.sh | sh Or paste that line to your AI and let it drive.

Early days: 0 graded parts across 0 jobs. Each bar averages every attempt on file, and the ticks are the attempts themselves; a small sample should look like one.

Grading is a fixed rubric measured on the part's actual geometry, so the only randomness is the model's. Raw results, full transcripts, and the grading code are on GitHub.