NVIDIA Releases Kumo Tabular: Open Tabular Foundation Models Claiming a New Accuracy-Efficiency Frontier

NVIDIA announced Kumo Tabular via its Hugging Face blog, claiming a new accuracy-efficiency frontier for tabular prediction; the family comprises Small, Medium, and Large models (about 28M–215M parameters) that predict new rows in a single forward pass, with…

NVIDIA first announced Kumo Tabular on the Hugging Face blog on 2026-09-29, claiming it sets a new accuracy-efficiency frontier for tabular prediction; no article text was available from that announcement, so it contributes only NVIDIA's framing. A MarkTechPost report on 2026-10-01 filled in the details: Kumo Tabular is a family of tabular foundation models in Small, Medium, and Large sizes spanning roughly 28 million to 215 million parameters. The models classify or regress new rows in a single forward pass from labeled context, requiring no per-dataset training, hyperparameter tuning, or feature engineering. Weights ship under OpenMDW-1.1, which permits commercial use (unlike several rivals), and run through the Apache-2.0 structured-data-models library. NVIDIA reports first place on TabArena with an Elo of 1950 and 17x faster inference than LimiX-2 on a single RTX 6000 Pro. Pretraining used only synthetic structural causal model tables of up to 60,000 rows, and regression outputs 999 quantiles as an uncertainty estimate. The two sources are consistent; the Hugging Face post simply adds NVIDIA's own frontier claim without further technical detail.

• Three model sizes — Small, Medium, and Large — span approximately 28 million to 215 million parameters.
• Models predict new rows via classification or regression in a single forward pass from labeled context, with no per-dataset training or feature engineering.
• Weights are released under OpenMDW-1.1, which permits commercial use, unlike several rivals; inference runs through the Apache-2.0 structured-data-models library.
• NVIDIA reports first place on TabArena with an Elo of 1950.
• NVIDIA reports 17x faster inference than LimiX-2 on one RTX 6000 Pro.
• Pretraining used only synthetic structural causal model tables, up to 60,000 rows.
• Regression outputs 999 quantiles as an uncertainty estimate.
• NVIDIA's Hugging Face blog announcement (2026-09-29) claims Kumo Tabular sets a new accuracy-efficiency frontier for tabular prediction; no article text was available from that post.