NVIDIA releases Kumo Tabular, an open model that predicts from tables without training
NVIDIA released Kumo Tabular, an open foundation model for tabular data. It uses labeled rows as examples and predicts new rows for classification and regression in a single pass, with no training, tuning or feature engineering. It comes under a license that allows commercial use and can run on your
NVIDIA announced on Hugging Face on September 29, 2026 Kumo Tabular, an open foundation model for tabular data. The model reads a table of labeled rows as context and predicts the labels of new rows in a single forward pass. No training, tuning or feature engineering is needed, and it works for both classification and regression.
Predictions straight from a table
Much of a company's data sits in tables such as customer records, transactions, sensor logs and orders, and churn, default risk, demand or price forecasts are built from them. According to NVIDIA's post, each new question has meant collecting labels, preparing features, searching settings and training a separate model. With Kumo Tabular, rows where the target column is filled in serve as examples, and the model returns results for the empty rows directly: class probabilities for classification and numeric predictions for regression.
The model works with numerical and categorical columns; text, images or timestamps can be turned into features with built-in preprocessing recipes in the library. Missing values need no imputation, as the model handles them on its own. A single pass covers up to 10 classes, and the library handles more classes with error-correcting output codes. The model was trained only on artificially generated tables, and the last stage of training used tables of up to 60,000 rows and 100 columns.
Sizes, license and setup
Kumo Tabular comes in three sizes, from 28 million to 215 million parameters. The model is released under the OpenMDW-1.1 license and can be used commercially. It runs through NVIDIA's open-source structured-data-models library, which installs with pip install structured-data-models and downloads the model files from Hugging Face on first use. A pandas table and a few lines of Python are enough to get predictions. The library is built for GPUs: the examples run the model on CUDA and require Python 3.11 and PyTorch 2.7 or later. NVIDIA says it ran its benchmarks on a single RTX 6000 Pro card; the pages do not list minimum hardware requirements.
According to the company, the model ranks first on the TabArena, BeyondArena, TALENT and ScoringBench tabular benchmarks. The training recipe and the artificial data generators are said to be coming soon.
Because it is a model you can run on your own servers, tasks such as predicting churn from CRM records, stock needs or equipment failures can be done without data leaving the company. With no separate model to train for each question, it is also an easy option to try for companies with small data science teams.
NVIDIA writes that accuracy may drop on tables far beyond the training range or when the rows to be predicted come from a different distribution than the example rows, and it recommends checking results on the company's own held-out data before deployment. Since the model was trained only on artificial tables, its results on real business data may vary by task. To see how the benchmark results hold up on other datasets, it is worth following evaluations from independent researchers.
Sources: Hugging Face: NVIDIA Kumo Tabular announcement, Hugging Face: nvidia/Kumo-Tabular, GitHub: NVIDIA/structured-data-models.
