Kumo Tabular

Kumo Tabular is an AI model that makes predictions from a table. Give it rows with known answers and new rows with missing answers, and it predicts a category or a number. Nvidia released it on 29 September 2026 as part of its Kumo Structured model collection.

It learns the task from example rows

Kumo Tabular was trained on artificial data. At use time, it can infer a new table’s task from labelled examples, without a separate training run for that dataset. Classification chooses a category; regression estimates a numerical value. These are the two supported task types.

Nvidia offers three sizes spanning 28 million to 215 million parameters. Its launch report places the models first on four tabular benchmarks, using the company’s evaluation setup. A benchmark ranking describes the tested datasets and comparisons; users should measure accuracy on their own held-out rows.

The weights and supporting library are available

The model uses the OpenMDW 1.1 licence, which allows commercial use under its terms. Nvidia’s structured-data model library provides the implementation and usage documentation alongside other model families.

Kumo Tabular works on structured rows and columns. The practical decision is whether its predictions are accurate, fast and affordable on the table in question. Keeping test rows separate from the supplied examples is essential to measuring that fairly.