NVIDIA has put out Kumo Tabular, an open foundation model built for the tables that carry most enterprise machine learning.
The selling point is how little setup it needs. Point it at a table where some rows already carry labels and it returns predictions for the rest in one pass, covering classification and regression without separate models. There is no training run to launch, no hyperparameter search and no feature engineering, the rituals that have defined tabular work since gradient-boosted trees took over.
Training used nothing but artificial data, the company says. Weights are served through an open-source library rather than a hosted endpoint, and anyone shipping a commercial product falls under the OpenMDW-1.1 license. Three versions exist, spanning 28M to 215M parameters.
Rather than publish a single number, NVIDIA points to four separate leaderboards where it says the model places first: TALENT, ScoringBench, BeyondArena and TabArena. Kumo Tabular belongs to the wider Kumo Structured collection and can be downloaded from Hugging Face.
Predictions of this kind cover churn, default, demand and pricing drawn from customer records, transactions, claims and orders. Today, each fresh question means gathering labels and rebuilding a model from nothing. A pretrained tabular model aims to compress that cycle into one inference call.
NVIDIA is candid about the limits. Artificial pretraining can leave blind spots on unusual schemas, and the model still wants labeled examples from the target task before it earns its keep.
