NVIDIA's open table-prediction models top eight benchmarks, often by thin margins

Kumo Tabular predicts on spreadsheet-style data without training on it. It ranks first on eight public tests and is fast, but several leads sit within statistical noise.

5 min readIIIOpen sourcePreprint, not peer reviewed
Save
Eight parallel wooden rulers rest on off-white paper, each carrying a grey bead and a slightly leading blue bead.AI illustration
Illustration: AI generated

What the study shows

The paper describes Kumo Tabular, a family of three tabular foundation models. These are models pretrained once that can then predict on a new table straight away. They read the rows with known answers as examples and fill in the rows without answers. No training or tuning on the new table is needed. The three sizes have about 27 million, 61 million and 214 million parameters, the adjustable numbers inside a model.

The models learned only from artificial tables. A generator draws random cause-and-effect structures and turns them into tables. It adds the mess of real data: noisy measurements, groups of related columns, skewed targets and missing values.

The authors report that the largest model, Kumo Tabular-L, ranks first on eight benchmarks. These cover broad sets of real tables, probabilistic forecasts, tables with text or images, and very wide scientific tables. Most results come from public leaderboards. The authors scored two of the eight themselves.

Speed is the other claim. On TabArena, a widely used benchmark, Kumo Tabular-L predicts about 36 times faster than the runner-up, LimiX-2, timed on the same GPUs. On BeyondArena it is about 1.7 times faster than TabPFN-3.5. Accuracy rose with model size on every benchmark where sizes were compared.

The authors also name weak spots. On data split by group, such as unseen customers, it ranks fourth. On tables with more than 100,000 training rows it ranks ninth, behind several tuned tree and neural network methods. This is a preprint and has not been peer reviewed.

Our reading

In our reading, the headline "first on eight benchmarks" needs care. The paper itself is candid here. On all six leaderboards that use Elo ratings, a score from head-to-head comparisons, the uncertainty range of Kumo Tabular-L overlaps the best rival's. On TabArena it wins 49% of comparisons against LimiX-2. That is a tie, not a win.

Where the evidence is firmer, it is worth noting. Against the best rival, Kumo Tabular-L wins on clearly more datasets on four benchmarks, including 95 of 108 tables on STRABLE. The gap to traditional methods is wide. On TabArena, the best tuned and combined tree models rank well below every Kumo size.

On our readiness scale, this work is at level III, open source available. Code is released under Apache 2.0 and weights under OpenMDW-1.1, though we could not open the links ourselves. It is not at level IV: the paper documents no use in production by a named organisation.

Free membership

Read the rest of this story

Membership is free. You get every article in full and the Issue every Sunday.

We send a sign-in link, no password needed. Already a member? Sign in

Readiness assessment

III
Open source
LevelStatusEvidence
I  Lab onlyMetWorking models with an architecture, pretraining recipe and ablations are described in the paper.
II  BenchmarkedMetPublic benchmark results on eight leaderboards (TabArena, TALENT, BeyondArena, ScoringBench, STRABLE, MulTaBench, RamanBench, TabBench-Bio). Kumo Tabular-L is ranked first on each, with confidence intervals shown on the Elo boards.
III  Open sourceMetPretraining, inference and prior code is released under Apache-2.0 at github.com/NVIDIA/structured-data-models. Weights for classifier and regressor in all three sizes are at huggingface.co/nvidia/Kumo-Tabular under OpenMDW-1.1. We could not open the links ourselves.
IV  In productionNot yetThe paper documents no deployment by a named organisation. The benchmark maintainers are thanked for listing it, which is not production use.

What would move it up
To reach level 4, a named organisation would need to document production use, for example on its own tabular prediction task with latency, cost and accuracy compared against tuned tree models. Independent replication of the benchmark wins would also help. So would results on grouped, temporal and very large tables, where it is weaker.

The study

Jingang Qu, Valter Hudovernik, Martin Jurkovič et al.. NVIDIA Kumo Tabular: Pushing the Accuracy–Efficiency Frontier for Tabular Prediction. arXiv, 17 November 2026. Read the paper

Preprint, not peer reviewed · Licence arXiv (no reuse)

Notes

  1. Qu, J. et al. NVIDIA Kumo Tabular: Pushing the Accuracy-Efficiency Frontier for Tabular Prediction. arXiv preprint, 17 November 2026, not peer reviewed. Source
  2. Pretraining, inference and data-generator code, Apache 2.0 licence. Source
  3. Model weights, classifier and regressor in three sizes, OpenMDW-1.1 licence. Source
  4. TabArena Elo: Kumo Tabular-L 1962 (95% interval 1862 to 2124) against LimiX-2 1952 (1866 to 2082); 49% head-to-head win rate.
  5. Head-to-head wins against the best non-Kumo entry: TALENT 132 to 95, ScoringBench 68 to 33, TabBench-Bio 32 to 9, STRABLE 95 to 13, MulTaBench 40 to 40.
  6. Training compute per checkpoint: about 800, 3,600 and 11,600 GPU-hours on NVIDIA H100 GPUs for the S, M and L sizes.

Edited and signed by

Stefan PerdalEditor and publisher86 stories edited

Stefan read the study, set the readiness level and approved this story. An AI system found the paper and wrote the first draft.

How this was made

An AI model read the study and wrote this article in the structure Science Underway uses. An editor reviewed it and approved it for publication. Check the original study before you rely on the details.

More on Efficiency