Breaking

NVIDIA releases Kumo Tabular, an open model that scores tables without training

September 29, 2026 at 17:51 EDT

Foundation Models
Open Source

NVIDIA on Tuesday released Kumo Tabular, an open family of foundation models that predict the labels of new table rows in a single forward pass, without training, tuning, or feature engineering for each task. The models are part of the NVIDIA Kumo structured-data collection, and the company is shipping open weights plus an open-source inference library under terms that allow commercial use.

September 29, 2026 · NVIDIA

Table labels, predicted in one pass

Kumo Tabular is an open foundation model that reads labeled rows and scores the rest without training, tuning, or feature engineering.

Relative inference speed

17× versus LimiX-2 on one GPU

Block count equals the reported speed. NVIDIA timed both on a single RTX 6000 Pro.

1×

LimiX-2

17×

Kumo Tabular

The taller stack is reported speed, not a larger parameter count.

1950

TabArena Elo, reported first overall

1418

BeyondArena Elo, also reported first

28–215M

Parameters across small, medium, and large

7.78%

BeyondArena Improvability score

Pretraining volume

Artificial tables, by model size

Stacked height is proportional to 35 million, 71 million, and 137 million synthetic tables.

35M

Small

71M

Medium

137M

Large

How a prediction is made

1 · Labeled rows

Features and known targets become the context.

→

2 · One forward pass

No weight update and no hyperparameter search.

→

3 · Scores

Class probabilities, or 999 regression quantiles.

What users can use now

Open weights, Apache-2.0 library code, and a license that allows commercial use. No per-task training. Reported first place on four tabular benchmarks, plus a 17× speed claim against LimiX-2.

What is still unsettled

The ranks are NVIDIA's own runs. Native columns are numerical or categorical. More than 10 classes need extra coding. Accuracy can slip if new rows differ from the context, and the training generators are not released yet.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.