Skip to main content

NVIDIASpecialist modelOpen weights

Kumo-Tabular

PopularityRank 4 of 19 specialist models

45out of 100new calculation

based on two data groups, mostly search, range 35 to 54

Summary

Kumo-Tabular is NVIDIA’s tabular foundation model for classification and regression. You provide labeled context examples and query rows, and the model returns class probabilities or regression values. It works on structured tables and is not a System One chat model.

NVIDIA publishes three sizes with 28-215 million parameters under OpenMDW 1.1. The weights run locally through structured-data-models. No Hugging Face Inference Provider or specific context window is documented.

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 60 published benchmark results

Page 1 of 5

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

BeyondArena Improvability score (Show measurement, test conditions, and source)
Source value
7.78
Metric
Improvability score
Unit
%
Direction
No direction published
Source type
vendor-reported
Evaluator
NVIDIA
Status
active
Retrieved at
2026-10-18
Methodology
Vendor-reported result. The report names Kumo Tabular. Size/checkpoint and benchmark-specific settings are not stated for these results; no additional configuration is inferred.
Source NVIDIA
TabArena overall Elo, Kumo-Tabular default/Large (Show measurement, test conditions, and source)
Source value
1,990
Metric
overall Elo rating
Unit
elo
Benchmark version
TabArena-v0.1.9.6
Category
tabular
Direction
Higher is better
Lower confidence bound
1888
Upper confidence bound
2134
Source type
official-leaderboard
Evaluator
TabArena official leaderboard
Status
superseded
Retrieved at
2026-10-15
Methodology
Official benchmark leaderboard. TabArena-v0.1.9.6; verified Kumo-Tabular (default) row in the all-task, all-dataset, no-imputation, all-splits leaderboard. NVIDIA's released config sets default_size=large; Large is approximately 215M parameters. This is the default/Large checkpoint row, not a family aggregate.
Context
TabArena Elo is a relative rating for this leaderboard population, not a percentage. The CSV reports an asymmetric 95% confidence interval of +144/-102 Elo.
Source TabArena official leaderboard
TabArena overall Elo, Kumo-Tabular default/Large (Show measurement, test conditions, and source)
Source value
1,992
Metric
overall Elo rating
Unit
elo
Benchmark version
TabArena-v0.1.9.7
Category
tabular
Direction
Higher is better
Lower confidence bound
1891
Upper confidence bound
2137
Source type
official-leaderboard
Evaluator
TabArena official leaderboard
Status
superseded
Retrieved at
2026-10-15
Methodology
Official benchmark leaderboard. TabArena-v0.1.9.7; verified Kumo-Tabular (default) row in the all-task, all-dataset, no-imputation, all-splits leaderboard. NVIDIA's released config sets default_size=large; Large is approximately 215M parameters. This is the default/Large checkpoint row, not a family aggregate.
Context
TabArena Elo is a relative rating for this leaderboard population, not a percentage. The CSV reports an asymmetric 95% confidence interval of +145/-101 Elo. The earlier v0.1.9.6 observation is retained separately; changes in the leaderboard population do not prove a checkpoint improvement.
Source TabArena official leaderboard
TabArena overall Elo, Kumo-Tabular Medium (Show measurement, test conditions, and source)
Source value
1,930
Metric
overall Elo rating
Unit
elo
Benchmark version
TabArena-v0.1.9.7; OpenMDW-1.1; leaderboard revision 0bc60d4d8838afc8cdcd5d05550851d39af54d01
Direction
Higher is better
Lower confidence bound
1852
Upper confidence bound
2033
Source type
official-leaderboard
Evaluator
TabArena official leaderboard
Status
superseded
Retrieved at
2026-10-16
Methodology
Official benchmark leaderboard. Kumo-Tabular Medium checkpoint; exact NVIDIA checkpoint config revision c5893d418e162b9f841ae179dc767afefff73831e; OpenMDW-1.1; TabArena-v0.1.9.7; leaderboard-reported 95% Elo interval +103/-78.
Context
Medium is a separate checkpoint, not the default/Large model. CI is the leaderboard's reported Elo interval, not a score range across tasks.
Source TabArena official leaderboard
TabArena overall Elo, Kumo-Tabular Small (Show measurement, test conditions, and source)
Source value
1,824
Metric
overall Elo rating
Unit
elo
Benchmark version
TabArena-v0.1.9.7; OpenMDW-1.1; leaderboard revision 0bc60d4d8838afc8cdcd5d05550851d39af54d01
Direction
Higher is better
Lower confidence bound
1764
Upper confidence bound
1905
Source type
official-leaderboard
Evaluator
TabArena official leaderboard
Status
superseded
Retrieved at
2026-10-16
Methodology
Official benchmark leaderboard. Kumo-Tabular Small checkpoint; exact NVIDIA checkpoint config revision c5893d418e162b9f841ae179dc767afefff73831e; OpenMDW-1.1; TabArena-v0.1.9.7; leaderboard-reported 95% Elo interval +81/-60.
Context
Small is a separate checkpoint, not the default/Large model. CI is the leaderboard's reported Elo interval, not a score range across tasks.
Source TabArena official leaderboard
AIM-Tabular (Show measurement, test conditions, and source)
Source value
1,911
Metric
AIM-Tabular Elo rating
Unit
elo
Benchmark version
AIM-Tabular; AIMultiple page snapshot 2026-10-17; Kumo added 2026-10-12
Category
tabular prediction
Direction
Higher is better
Source type
independent-evaluator
Evaluator
AIMultiple AIM-Tabular Index
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. The exact source row is Kumo-Tabular. AIM-Tabular uses 13 models across 19 datasets, 5-fold cross-validation, default hyperparameters, and one NVIDIA B200 for foundation models. NVIDIA config commit c5893d418e162b9f841ae179dc767afefff73831 sets the default to the Large checkpoint and maps Small, Medium, and Large separately. AIMultiple does not publish the tested weight-file revision.
Context
The displayed 95% interval belongs to the combined index, not the AIM-Tabular cell. The source does not give a per-cell sample N or confidence interval, and does not expose the exact evaluated weight revision.
Source AIMultiple AIM-Tabular Index
BeyondArena ELO (Show measurement, test conditions, and source)
Source value
1,418
Metric
ELO
Unit
elo
Direction
No direction published
Source type
vendor-reported
Evaluator
NVIDIA
Status
active
Retrieved at
2026-10-18
Methodology
Vendor-reported result. The report names Kumo Tabular. Size/checkpoint and benchmark-specific settings are not stated for these results; no additional configuration is inferred.
Source NVIDIA
TALENT average rank, classification accuracy (Show measurement, test conditions, and source)
Source value
6.67
Metric
Average rank across classification accuracy
Unit
No unit provided
Direction
No direction published
Source type
vendor-reported
Evaluator
NVIDIA
Status
active
Retrieved at
2026-10-18
Methodology
Vendor-reported result. The report names Kumo Tabular. Size/checkpoint and benchmark-specific settings are not stated for these results; no additional configuration is inferred.
Source NVIDIA
TALENT average rank, classification log-loss (Show measurement, test conditions, and source)
Source value
3.98
Metric
Average rank across classification log-loss
Unit
No unit provided
Direction
No direction published
Source type
vendor-reported
Evaluator
NVIDIA
Status
active
Retrieved at
2026-10-18
Methodology
Vendor-reported result. The report names Kumo Tabular. Size/checkpoint and benchmark-specific settings are not stated for these results; no additional configuration is inferred.
Source NVIDIA
TALENT average rank, regression RMSE (Show measurement, test conditions, and source)
Source value
4.22
Metric
Average rank across regression RMSE
Unit
No unit provided
Direction
No direction published
Source type
vendor-reported
Evaluator
NVIDIA
Status
active
Retrieved at
2026-10-18
Methodology
Vendor-reported result. The report names Kumo Tabular. Size/checkpoint and benchmark-specific settings are not stated for these results; no additional configuration is inferred.
Source NVIDIA
AIM Tabular Index (Show measurement, test conditions, and source)
Source value
1,692
Metric
combined Elo
Unit
elo
Category
tabular
Direction
Higher is better
Source type
independent-evaluator
Evaluator
AIMultiple
Status
active
Retrieved at
2026-10-18
Methodology
Independent evaluation. AIMultiple combines its AIM-Tabular benchmark with public TabArena and BeyondArena results. Its method shifts each benchmark so default XGBoost scores 1000, then averages the shifted ratings. The row displays an AIM Tabular Index score of 1692 beside the published index range 1610-1794. The source does not explicitly identify this composite range as a 95% confidence interval. The methodology states that each benchmark gives every model a 95% confidence interval and that these are averaged in the same way as scores. Public benchmarks contribute default-configuration scores and AIM-Tabular runs every model without tuning or fine-tuning. AIM-Tabular is described as 13 models, 19 datasets, five-fold cross-validation and default hyperparameters; TabArena as 51 datasets; BeyondArena as 142 datasets, 49 of which overlap TabArena. These dataset counts are not a single evaluated N for the composite score.
Context
This is the evaluator's published composite Elo index, not a newly calculated score. Preserve 1610-1794 as the source-published AIM Tabular Index range; the methodology discusses 95% confidence intervals for each benchmark and averaging them but does not explicitly identify the displayed composite range as a 95% confidence interval. The baseline already contains the component AIM-Tabular, TabArena and BeyondArena measurements; this composite has no benchmark-specific profile comparison or rank authorization.
Source AIMultiple
Texas crash-table reader: mean flag AUC (small) (Show measurement, test conditions, and source)
Source value
0.865
Metric
Mean flag AUC across eight flag targets
Unit
No unit provided
Benchmark version
Texas CRIS 2017-2025; arXiv:2610.10321 (2026-10-07)
Category
tabular-prediction
Direction
Higher is better
Sample
30000
Source type
independent-evaluator
Evaluator
Amir Rafe and Subasish Das, Estimating Uncoded Crash Factors with Tabular Foundation and System One Models: Kumo Tabular and Jev (arXiv:2610.10321)
Status
active
Retrieved at
2026-10-19
Methodology
Independent evaluation. Kumo Tabular small; Table 6; mean flag AUC; same 30,000 test rows from one wave-1 fold; nested contexts drawn from other folds; context 60000 rows.
Context
Independent, task-specific evaluation on a Texas crash table, not a general tabular leaderboard. Table 6 reports one wave-1 test fold and no confidence interval or evaluated Kumo weight revision.
Source Amir Rafe and Subasish Das, Estimating Uncoded Crash Factors with Tabular Foundation and System One Models: Kumo Tabular and Jev (arXiv:2610.10321)

Page 1 of 5

Popularity compared

How often the model is searched for, used and mentioned, compared with every model of its category.

Popularity index

0 to 100, higher is more popular

20 of 21 models

As of November 2026 · Source: gradually.ai popularity index

Over time

Weekly values from October 18, 2026 to November 20, 2026

Kumo-TabularUnlimited OCR 2 55TeleOCR 38Agora-2 33Qwen-Drive-1.0-4B 29

Method changes in the timeline, a jump at these points is no market event:

  1. 1November 1, 2026: Value only with search or usage data, 30-day attention
  2. 2November 1, 2026: Method version 4 (common scale, shrinkage)
  3. 3November 9, 2026: New topics measurement V20 (daily quorum, short forms)

What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.

All specifications

Architecture

Model class
Tabular foundation model
Source: Model class
Parameters
28-215 million parameters across three sizes (Small, Medium, and Large)
Source: Parameters

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking
Unknown
no source
Watermarking technique
Not known
no source

Licence and access

Access
Local execution through structured-data-models
Source: Access
License
OpenMDW 1.1
Source: License
Shutdown
None announced
no source

API

Model ID
nvidia/Kumo-Tabular
Source: Model ID

Further details

Tasks
Classification and regression with labeled context examples
Source: Tasks
Hosted service
No Inference Provider on Hugging Face
Source: Hosted service

Sources and data date

Show all 33 sources

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.

Without vendor claims and historical or disputed results; those stay labelled in the single values.

TypeEvidence and data date
Research statusResearch date October 4, 2026. 27 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not published

aim-tabular-cell-confidence-interval

Original wording

The page gives an interval for the combined index only; no interval is published for the separate AIM-Tabular rating.

Not published

evaluated-weight-revision

Original wording

AIMultiple names Kumo-Tabular and default hyperparameters but does not report a weight-file revision. The public NVIDIA default configuration resolves the default size to Large.

Unresolved

checkpoint-and-protocol

Original wording

For the BeyondArena and TALENT results, the inspected article does not specify evaluated size/checkpoint, benchmark-specific settings, direction, sample size, or confidence interval.

Unresolved

checkpoint

Original wording

The leaderboard identifies the model as Kumo-Tabular NVIDIA but does not state a checkpoint or model revision.

Not published

evaluated sample count

Original wording

The page lists dataset counts for component benchmarks but does not publish one evaluated sample count for the composite index; dataset counts are not treated as completed model cases.

Not published

Time-series forecasting benchmark score

Note

NVIDIA lists forecasting from tabular features as a use case. These TabArena/BeyondArena evaluations cover tabular classification/regression and exclude dedicated time-series forecasting applications.

Additional sourceNVIDIA Kumo-Tabular model card (retrieved October 3, 2026) · NVIDIA Kumo-Tabular model card · Editorial description reviewed October 3, 2026
Additional sourceNVIDIA Kumo-Tabular announcement (retrieved October 3, 2026) · NVIDIA Kumo-Tabular announcement · Editorial description reviewed October 3, 2026
Additional sourceTabArena official leaderboard (retrieved October 15, 2026)
Additional sourceTabArena official leaderboard (retrieved October 15, 2026; October 16, 2026)
Additional sourceAIMultiple AIM-Tabular Index (retrieved October 17, 2026; October 18, 2026)
Additional sourceNVIDIA (retrieved October 18, 2026)
Additional sourceAmir Rafe and Subasish Das, Estimating Uncoded Crash Factors with Tabular Foundation and System One Models: Kumo Tabular and Jev (arXiv:2610.10321) (retrieved October 19, 2026)
Additional sourceMulTaBench official leaderboard (retrieved October 20, 2026)
Additional sourceCyberNative AI LLC, Kumo Label Lens v2 (retrieved October 22, 2026)
Additional sourceTabArena official leaderboard (entrants=models) (retrieved October 28, 2026)
Additional sourceTabArena official leaderboard (entrants=open) (retrieved October 28, 2026)
Additional sourceTabArena BeyondArena official leaderboard (split=all) (retrieved October 28, 2026)
Additional sourceTabArena BeyondArena official leaderboard (split=random) (retrieved October 28, 2026)
Additional sourceTabArena BeyondArena official leaderboard (split=temporal) (retrieved October 28, 2026)
Additional sourceTabArena BeyondArena official leaderboard (split=grouped) (retrieved October 28, 2026)
Additional sourceModel metadata (retrieved October 15, 2026; October 28, 2026)
Additional sourceModel metadata (retrieved October 15, 2026)12
Additional sourceModel metadata (retrieved October 16, 2026) (unavailable on November 2, 2026) https://huggingface.co/nvidia/Kumo-Tabular/blob/c5893d418e162b9f841ae179dc767afefff73831e/config.json
Additional sourceModel metadata (retrieved October 17, 2026)
Additional sourceModel metadata (retrieved October 20, 2026)12
Additional sourceModel metadata (retrieved October 28, 2026)12345
Additional sourceOpenRouter: Popularity index (derived), CC BY 4.0
Additional sourceVercel AI Gateway: Popularity index (derived), CC BY 4.0
Additional sourceOllama: cumulative downloads for the popularity index
Additional sourceReplicate: cumulative runs for the popularity index
Additional sourceGoogle Trending Now: limited feed selection for Germany and the US
Additional sourceCreative Commons Attribution 4.0