The news
Nvidia (NVDA) on September 29 released Nvidia Kumo Tabular, an open foundation model that predicts outcomes from rows and columns of business data without training, tuning or feature engineering. Nvidia says the model ranks first on four public benchmarks.
The model handles classification, such as sorting customers into those likely to stay or leave, and regression, which predicts a number such as a price. It relies on in-context learning: a user passes in example rows with known answers alongside the rows to be predicted, and the model returns predictions in a single forward pass, with no separate fitting step. Nvidia said it was pretrained only on artificially generated tables built from structural causal models, which simulate cause-and-effect links between columns.
Kumo Tabular comes in three sizes, from 28 million to 215 million parameters. According to Nvidia, the Small, Medium and Large versions saw about 35 million, 71 million and 137 million artificial tables in pretraining. The final training stage used tables of up to 60,000 rows and 100 columns.
Nvidia reported an Elo rating of 1950 on TabArena, placing first overall in a field that included tuned gradient-boosted trees, AutoGluon and other tabular foundation models, while running 17 times faster than the rival LimiX-2 model on a single RTX 6000 Pro GPU. It also reported first place on BeyondArena with an Elo of 1418, the top overall ranking on TALENT, and first and second place for its Large and Medium versions on ScoringBench. Nvidia said the model sets a new state of the art for accuracy relative to compute cost.
The weights are on Hugging Face under the OpenMDW-1.1 license, which Nvidia says permits commercial use. The code sits in Nvidia's open-source structured-data-models library on GitHub, which recommends a GPU and also includes Kumo Relational, a model for predictions across linked database tables.
Nvidia listed limits. The model reads only numerical and categorical columns, so text, images or timestamps must first be converted with built-in preprocessing recipes. One pass covers up to 10 classes, and accuracy may degrade on tables unlike its training data. The benchmark figures are Nvidia's own: a pull request to add Kumo Tabular to TabArena, opened September 28 in the project's GitHub repository, said leaderboard positions would be posted once maintainer runs finished.
The numbers
- TabArena Elo (Nvidia-reported)
- 1950, first overall
- Speed vs. LimiX-2 (Nvidia test, one RTX 6000 Pro)
- 17x faster
- BeyondArena Elo (Nvidia-reported)
- 1418, first
- Model sizes
- 28M to 215M parameters
- Artificial pretraining tables (Small / Medium / Large)
- About 35M / 71M / 137M
- Largest pretraining context
- 60,000 rows, 100 columns
Why CEOs should care
For operations and analytics leaders, the pitch is speed to a first answer. Nvidia notes that customer records, transactions, claims and orders all live in tables, and that predicting churn, default, demand or price from them is among the most common machine learning jobs in industry. A model that predicts without a training run lets a team try a new forecasting question before building anything. The question for your data team: does Kumo Tabular beat our current models on our own held-out data, not just on public benchmark averages?
For CFOs and CIOs, the cost comparison matters as much as accuracy. Nvidia's library recommends a GPU, and its speed figures were measured on an RTX 6000 Pro card. Before switching any workload, compare the full cost of GPU inference with your current setup, and have legal review the OpenMDW-1.1 terms even though Nvidia says they allow commercial use.
For CISOs and risk officers, open weights mean the model can run inside your own environment, so labeled customer rows need not leave it. But in-context learning feeds those rows into the model at every prediction, which makes data handling part of the inference pipeline. Teams that use predictions in lending, insurance or hiring should also ask whether they can explain each result well enough for their own review and audit requirements.
The bigger picture
Kumo Tabular is part of a wider Nvidia push into software for structured business data, not only chips. Barchart reported in a June 5 article on Yahoo Finance that Nvidia had acquired Kumo AI, developer of the KumoRFM relational foundation model, at a reported value of $400 million. Nvidia now offers Kumo Relational through its NIM inference service.
The field is crowded. For two decades, Nvidia notes, this kind of prediction has been done with gradient-boosted trees, and its post names newer rivals including TabPFN, TabICL, LimiX-2 and AutoGluon. Kumo Tabular borrows column, row and in-context attention ideas introduced in TabICL and TabPFN, Nvidia said. With several teams building on similar ideas, benchmark leadership can change hands quickly.
What’s next
Watch for TabArena's own results from the maintainers' runs, which will show whether Nvidia's first-place ranking holds under independent evaluation. Companies with many tree-based models can download the weights and test Kumo Tabular on a few well-understood prediction tasks before committing any production workload.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
Companies in this story
