September 29, 2026
Tabular forecasts can run on a ready-made model: KumoTabular launched Sep 28
On Sep 28, 2026, NVIDIA released KumoTabular for classification and forecasting with tabular data. The lineup has three sizes: small, medium, and large, with large selected by default. Running it requires the `structured-data-models` package, Python 3.11+, and PyTorch 2.7+.
Without caching, the same context would have to be recomputed for each batch. KumoTabular retains these computations after `fit()`, then uses `predict()` for new batches.
Low barrier to entry. Install the package with `pip install structured-data-models`. The model supports classification and regression, and returns 999 quantiles for regression, from `q001` to `q999`.
The code is published under Apache-2.0, while KumoTabular weights are distributed under OpenMDW 1.1.
