NVIDIA Structured Data Models enable zero‑training tabular predictions

Zero‑training classification is demonstrated on the breast‑cancer dataset with three hundred labelled rows and two hundred and sixty‑nine rows to predict. TabICLv2 and KumoTabular achieve 0.970 accuracy and AUCs of 0.996 and 0.995, matching logistic regression (0.970 accuracy, 0.993 AUC) and out‑performing gradient boosting (0.967 accuracy, 0.992 AUC). Regression on the diabetes dataset returns 999 quantiles per query; the median prediction has an RMSE of 53.6, slightly better than RidgeCV (54.3) and far better than gradient boosting (64.7). Adding calendar and TFIDF preprocessing to TabICLv2 raises AUC from 0.525 to 0.789 on a support‑ticket escalation task. KumoRelational applied to a synthetic shop database (800 customers, 60 products, many orders) reaches 0.823 AUC, comparable to hand‑engineered features (0.828) but drops to 0.940 when time columns are omitted, exposing leakage.