NVIDIA Kumo Tabular Makes Predictions from Labeled Rows in One Pass

NVIDIA Kumo Tabular is a family of models for classification and regression that uses labeled rows as context to predict new rows in a single forward pass. The workflow avoids task-specific training, hyperparameter tuning and feature engineering, and comes in Small, Medium and Large sizes spanning about 28 million to 215 million parameters (NVIDIA Kumo Tabular release details).

That setup shifts effort away from fitting a new model for each table and toward choosing useful labeled context rows. It is an in-context prediction approach, not a claim that the model can make useful predictions without labeled examples.

Context rows do the work

Kumo Tabular uses separate models for classification and regression. Its prediction pipeline distinguishes context rows from query rows: context rows attend to each other, while query rows attend only to the context. The context representations can then be computed once and reused for queries (model architecture and outputs).

For classification, the model returns class probabilities. For regression, it outputs 999 quantiles, which the source describes as supporting a point prediction and an uncertainty estimate. NVIDIA says it pretrained the models entirely on synthetic tables generated from Structural Causal Models, with patterns such as missing values and high-cardinality categories included in the data (pretraining and prediction details).

Synthetic pretraining is relevant because the model is meant to transfer to tables it did not train on directly. But it does not, by itself, establish how well it will handle any particular organization’s data; the reported benchmark results offer one view of performance, not a guarantee for every dataset.

Deployment and licensing are separate questions

Kumo Tabular runs through NVIDIA’s open-source structured-data-models library, which uses a shared in-context learning interface based on TableTensor. The release account lists Python 3.11 or later and PyTorch 2.7 or later, with examples targeting a CUDA GPU (SDM library and requirements). The GPU requirement makes deployment environment a practical consideration, even though the prediction workflow avoids task-specific model training.

NVIDIA’s weights use the OpenMDW-1.1 license, which the source says permits commercial use; the SDM code is separately licensed under Apache-2.0. Those are distinct terms for the model weights and library, respectively (license details).

Benchmark leads are claims with a defined scope

The release account reports a first-place TabArena result with an Elo of 1950, and leading positions on BeyondArena, TALENT and ScoringBench. It also reports that Kumo Tabular ran 17 times faster than LimiX-2 on a single RTX 6000 Pro (reported benchmark results and hardware comparison). These are NVIDIA team results as reported in the article, not a universal speed or accuracy guarantee. The hardware detail matters: the 17x figure is a specific comparison, not a general measure of workload cost.

The strongest practical distinction is therefore the combination of single-pass prediction and commercial-use weights, alongside a GPU-oriented library. Whether that makes Kumo Tabular a fit depends on the table, the available deployment hardware and the task’s tolerance for uncertainty, not just a leaderboard position.