The Pulse
NVIDIA Releases Kumo Tabular With 215M Parameters
NVIDIA released Kumo Tabular, an open model for classification and regression on structured data, in three sizes from 28 million to 215 million parameters. The model uses labeled rows as context and makes predictions without task-specific t
AI.info Team · Written with AI ·
NVIDIA says its largest Kumo Tabular model learned from 137 million artificial tables before release on September 29, 2026. The new model family targets classification and regression: users provide labeled examples in a table, and the model predicts labels or values for new rows without task-specific training, tuning or feature engineering.
Three model sizes, one in-context approach
Kumo Tabular comes in small, medium and large versions ranging from 28 million to 215 million parameters. NVIDIA says the models were pretrained exclusively on artificial data, generated from structural causal models: simulated causal graphs produce columns and targets, then the process adds missing values, correlated features, outliers and other irregularities. The company says the small, medium and large models saw about 35 million, 71 million and 137 million synthetic tables, respectively.
At prediction time, labeled rows serve as context for rows with unknown outcomes. The model’s Transformer processes numerical and categorical values through column and row attention, then relates the known labels to the rows awaiting predictions. NVIDIA describes the method as a single forward pass, rather than a pipeline that retrains or tunes a separate model for each task.
“Given a table of labeled rows, it predicts the labels of new rows in a single forward pass, with no training, no tuning, and no feature engineering, for both classification and regression.”
NVIDIA’s Kumo Tabular authors
NVIDIA reports top results across four benchmarks
NVIDIA reports that Kumo Tabular ranks first on TabArena, BeyondArena, TALENT and ScoringBench. On TabArena, the company gives the model an overall Elo score of 1,950; on BeyondArena, it reports an Elo of 1,418 and an Improvability score of 7.78%. The post also says Kumo Tabular leads TALENT’s overall rankings for classification accuracy, classification log-loss and regression RMSE, and that its large and medium versions take the top two average-rank positions on ScoringBench.
Those are NVIDIA’s own benchmark results, not an independent evaluation. The company says it ran the three sizes with default settings on TabArena, comparing them with tuned gradient-boosted trees, AutoGluon and other tabular foundation models. It says the evaluation used one RTX 6000 Pro setup, but does not provide a clear speed multiplier in the post’s text. NVIDIA says its training recipe and artificial-data generators will be released later, so outside researchers cannot yet reproduce the full pretraining process from the release materials described there.
Useful for tables, but not every kind of data
The model handles numerical and categorical columns directly. Text, images and timestamps require preprocessing to turn them into features; one forward pass supports up to 10 classes, while NVIDIA’s library uses error-correcting output codes to extend classification to more classes. The weights are available on Hugging Face under the OpenMDW 1.1 license, and NVIDIA points developers to its open-source structured-data-models library for inference.
NVIDIA also warns that accuracy may fall when a table extends far beyond the ranges represented in training or when prediction rows differ in distribution from the labeled context rows. The company advises users to check accuracy and calibration on held-out data before deployment. That caveat matters for organizations considering a pretrained alternative to building a separate model for each dataset: the model removes several setup steps, but its release does not remove the need to test it on the data and decisions at hand.
Training details still to come
Kumo Tabular’s synthetic-only pretraining separates its learned patterns from any claim that it trained on private enterprise tables. But the release leaves open practical questions about how its benchmark lead holds up on specific business datasets and how much time inference saves in real deployments. The first test developers can run is available now through NVIDIA’s model weights and inference library; the company’s promised data generator and training recipe are not yet part of the release.