# NVIDIA Kumo Tabular released on Hugging Face with three model sizes

> The open foundation model for tabular data now available on Hugging Face comes in Small, Medium and Large versions (28 M‑215 M parameters) and tops four major benchmarks.

Oossa · 2026-09-29 · https://oossa.com/en/nvidia-kumo-tabular-released-on-hugging-face-with-three-model-sizes

NVIDIA has published Kumo Tabular on Hugging Face. The model predicts labels for new rows directly from a labeled table, without any training or feature engineering. Three sizes are offered – Small (≈28 M parameters), Medium (≈71 M) and Large (≈215 M). In the TabArena leaderboard it leads with an Elo of 1950 and runs about 17 % faster than the previous best LimiX‑2 when evaluated on a single RTX 6000 Pro GPU. The same models also rank first on BeyondArena, TALENT and ScoringBench.

## The facts

- Three model sizes: 28 M, 71 M, 215 M parameters (Hugging Face blog, Sep 29 2026)
- Elo 1950 on TabArena; 17 % faster than LimiX‑2 on RTX 6000 Pro (same source)

## Why it matters

It lets enterprises replace custom tabular pipelines with a single pretrained model that is both more accurate and quicker to run.

## Sources & references

1. [NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction](https://huggingface.co/blog/nvidia/kumo-tabular) – Hugging Face, 2026-09-29

Last updated: 2026-09-29