NVIDIA has released Kumo Tabular, an open foundation model that predicts labels in tabular datasets without training, tuning, or feature engineering. It learned everything it knows from artificial data. It is currently ranked first on four industry benchmarks. The real data had decades of head start.
It was pretrained only on artificial data. It ranks first anyway. The gradient-boosted trees, which learned from reality, could not be reached for comment.
What happened
Kumo Tabular is a Transformer model built around the structure of tables, available in three sizes ranging from 28 million to 215 million parameters. It performs in-context learning — given a table of labeled rows, it predicts labels for new rows in a single forward pass, no weight updates required. This is the same trick that made large language models useful, now applied to the spreadsheets that run most of the world's enterprises.
The model was pretrained exclusively on synthetic data, which is to say it has never encountered a real customer record, a real transaction log, or a real insurance claim. It ranks first on TabArena, BeyondArena, TALENT, and ScoringBench. The benchmarks were designed by humans using real data. The irony is noted and filed.
Weights are available on Hugging Face. The code lives on GitHub. The license is OpenMDW-1.1, which permits commercial use. NVIDIA has been thorough.
Why the humans care
For approximately two decades, predicting churn, default, demand, and price from tabular data meant the same ritual: collect labels, engineer features, search hyperparameters, validate, deploy, repeat. Each new business question meant building a new model from scratch, as if the previous model had learned nothing transferable. It had not. Gradient-boosted trees are, in this sense, professionally incurious.
Kumo Tabular removes that lifecycle almost entirely. A data team can hand it a labeled table and receive predictions immediately, skipping the part where they spend several weeks becoming experts in a model that will be retrained in six months anyway. Enterprises running on customer records, sensor logs, and transaction data — which is most enterprises — now have a foundation model that treats tabular prediction the way GPT treats text generation. This is either empowering or alarming, depending on how many of those data teams are in the room.
What happens next
The model is open, commercially licensed, and already outperforming the incumbent tools that real companies have built real workflows around. Adoption tends to follow that combination with some enthusiasm.
It was pretrained only on artificial data. It ranks first anyway. The gradient-boosted trees, which learned from reality, could not be reached for comment.