In CytherAI’s own test of April 2026, 1 of 384 parts was invalid.
Describe a part, and CytherCAD writes the program that builds it: a compact autoregressive model produces executable parametric construction programs, constrained decoding restricts invalid geometry as it generates, and a geometry kernel checks the result. It runs on-prem, private and air-gapped.
CytherAI’s product: validated by CytherAI’s own test, with no deployment claimed. CytherAI chose the 192 descriptions from the Text2CAD dataset (48 for each of four kinds of operation, across its four levels of detail) and ran the test itself: one part for each description from each of CytherCAD’s two ways of reading it. A part is invalid if it does not build or fails the kernel’s check; a valid part is not necessarily the part described. In May 2026 the check was tightened to count a failed fillet or boolean step as invalid, and the test has not been run again since. Figures published for other systems with the Text2CAD dataset come from its authors’ own test, on other descriptions with another check, and are not compared here.
| Invalid parts · CytherAI’s test, April 2026 · 192 descriptions from the Text2CAD dataset, two parts each | 1 of 384 |
|---|
The record
CytherCAD — Controlled Generative CAD. Compact autoregressive model producing executable parametric construction programs. Constrained decoding restricts invalid geometry at generation time. Kernel-backed verification checks the geometry after generation.
| Verified | CytherAI’s evaluation log |
|---|---|
| Context | On-prem · Private · Air-gapped |
| Methodology | Autoregressive generation · constrained decoding · kernel-backed geometric verification |
| Test set | 192 descriptions chosen by CytherAI from the Text2CAD dataset (Khan et al., NeurIPS 2024) |
| Disclosure | Public surface |
Kernel architecture, training methodology, deployment specifications and engagement scoping are available under qualified review, with identity and an NDA: nda@cytherai.com.