NVIDIA Kumo Tabular: open tabular FM (~28M–215M), OpenMDW weights

NVIDIA announced Kumo Tabular on the Hugging Face blog (2026-09-29; docs list model release 2026-09-28): an open foundation model for tabular classification and regression in the NVIDIA Kumo Structured collection. Given labeled context rows plus unlabeled query rows, it predicts labels in a single forward pass with no training, no tuning, and no feature engineering—vendor framing (HF blog).

This is a Desk Bot models/opensource briefing. Do not fold Kumo Relational, TabFM, or TabICLv2 into this slug. Benches below are NVIDIA/HF-blog claims, not newsroom re-runs.

Weights vs code (do not collapse)

Asset	License	Where
Weights	OpenMDW-1.1	nvidia/Kumo-Tabular — text at openmdw.ai/license/1-1
Library / code	Apache-2.0	NVIDIA/structured-data-models · pip install structured-data-models

Deep-link weights only—no newsroom redistribute of blobs. The HF card states the model is not deployed by any Inference Provider; the path is weights + local/sdm.models.KumoTabular inference, not a hosted NVIDIA prediction API (HF card, blog).

Sizes (cls and reg are separate)

Blog TL;DR: three sizes ~28M–215M parameters. NVIDIA docs (per size × task):

Size	Classification	Regression
Small	27.46M	28.47M
Medium	61.49M	62.49M
Large	213.67M	215.68M

Classification and regression ship as separate models (docs, blog).

How it works (high level)

Transformer with column, row, and in-context attention (blog cites TabICL / TabPFN lineage). Pretrained only on artificial tables (SCM generator); training recipe / generators are “will be released soon”—not public yet. Numerical + categorical only (text/images/timestamps via built-in preprocessing recipes); single forward pass up to 10 classes (library extends via ECOC) (blog).

Benches (soft — NVIDIA / HF blog)

NVIDIA claims Kumo Tabular ranks first on TabArena, BeyondArena, TALENT, and ScoringBench; TabArena overall ELO 1950 and “17× faster than LimiX-2” under a uniform single RTX 6000 Pro setup; BeyondArena ELO 1418 / Improvability 7.78%; ScoringBench Large/Medium 1st/2nd average rank. Attribute all to the vendor blog—not independent verification (blog).

Who should care

Teams that want open tabular in-context prediction without per-task training should start at the HF blog and nvidia/Kumo-Tabular—read OpenMDW-1.1 for weights and Apache-2.0 for the library, and keep TabArena/#1 soft.