Nvidia's open Kumo Tabular predicts spreadsheet rows in one pass, trained only on synthetic tables

In a Hugging Face blog post, Nvidia authors introduce Kumo Tabular, an open foundation model for tabular prediction within the Kumo Structured collection. Given a table with labeled rows, it predicts labels for new rows in a single forward pass, for both classification and regression, without fine-tuning or feature engineering. That is the same in-context approach large language models use for text, applied to spreadsheets and databases.

The model comes in three sizes, from 28 million to 215 million parameters, with an open-source library and weights released under the OpenMDW-1.1 license, which permits commercial use.

Kumo Tabular was pretrained only on artificial data: tens of millions of tables generated from structural causal models, with context lengths staged up to 60,000 rows. Its architecture uses attention across cells, rows and in-context examples, following the TabICL and TabPFN lineage, plus a length-aware attention temperature that helps with large tables.

Nvidia says the model ranks first on four benchmarks, TabArena, BeyondArena, TALENT and ScoringBench, citing an example TabArena Elo of 1950. It also says Kumo Tabular runs 17 times faster than LimiX-2 on a uniform RTX 6000 Pro setup. These results come from the authors; independent reproduction will help confirm them.

Much of business data lives in tables, from sales and inventory to risk scoring, and traditional approaches require training a separate model and engineering features for each task. A model that predicts from examples in one pass could let analysts get baseline predictions quickly. Nvidia has published related enterprise work, such as an agentic cybersecurity system with Nemotron and codifying supply-chain expertise.

Training only on synthetic tables avoids privacy and licensing issues with real data, but raises the question of how well the model handles messy real-world tables with missing values and unusual distributions. Real-world case studies, independent benchmark runs and integration with data tools will show how useful Kumo Tabular is in practice. More in our research coverage and tools coverage.