NVIDIA releases Kumo Tabular, an open model that predicts table rows without training and tops four benchmarks

Predicting churn, default, demand, or price from a table has been the job of gradient-boosted trees for two decades, and each new question meant collecting labels, engineering features, and searching hyperparameters. NVIDIA's Kumo Tabular, released on September 29, 2026 on the Hugging Face blog, applies in-context learning to that job. You hand it labeled rows as context and unlabeled rows as queries, and it returns class probabilities or numeric predictions without updating a single weight.

What it does

The model is a Transformer built around the structure of a table. Column attention learns what a value means within its column, row attention learns how features interact, and a final stage lets query rows attend to the labeled context. Because the context never looks at the queries, its keys and values are computed once and reused for follow-up predictions. For regression the head returns 999 quantiles, which give both a point prediction and an uncertainty estimate.

Kumo Tabular was pretrained only on artificial tables sampled from structural causal models. The three sizes saw about 35, 71, and 137 million such tables, in three stages that grew the context from 1,024 rows to 60,000 rows, always with up to 100 columns. Classification and regression are separate models.

Three sizes, one license

The weights sit on Hugging Face under the OpenMDW-1.1 license, which permits commercial use. The library that runs them, structured-data-models, is under Apache 2.0 and installs with pip install structured-data-models.

• Sizes: Small, Medium, and Large, from 28 million to 215 million parameters.
• Classifier files of roughly 110, 246, and 855 megabytes, and regressor files of roughly 114, 250, and 863 megabytes, converted by us from the byte sizes listed on Hugging Face.
• Python 3.11 and PyTorch 2.7 or newer, with NVIDIA recommending cuDF to keep dataframe operations on the GPU.
• No inference provider on Hugging Face serves the model; you run it on your own CUDA GPU.

The numbers NVIDIA reports

According to NVIDIA's measurement, Kumo Tabular ranks first on TabArena with an ELO of 1950 while running 17 times faster than LimiX-2, on a single RTX 6000 Pro under a uniform evaluation setup. On BeyondArena it reaches an ELO of 1418 with an Improvability score of 7.78%, again first. On TALENT it takes the top overall ranking with average ranks of 6.67 for classification accuracy, 3.98 for classification log-loss, and 4.22 for regression RMSE. On ScoringBench, a benchmark for predictive distributions, the Large and Medium models place first and second.

These are the vendor's figures against the public leaderboards, and no independent replication existed when this was written. The comparison set includes tuned gradient-boosted trees, AutoGluon, and the recent tabular foundation models.

Limits to read first

Kumo Tabular works on numeric and categorical columns only; text, images, and timestamps have to be turned into features through the library's preprocessing recipes. A single forward pass covers up to 10 classes, and the library extends that with error-correcting output codes. NVIDIA warns that accuracy may degrade on tables far beyond the training ranges, or when the query rows come from a different distribution than the context rows, and asks you to validate on held-out data before deployment. The training recipe and the data generators are promised for later, without a date.