NVIDIA released Kumo Tabular, an open-weight tabular foundation model that predicts classification or regression targets from labeled rows in one forward pass without task-specific training or feature engineering. Three models span 28M to 215M parameters and were pretrained entirely on synthetic tables. NVIDIA reports first-place aggregate results on TabArena, BeyondArena, TALENT and ScoringBench. The weights are ungated under OpenMDW-1.1 and inference code is Apache-2.0, but the training recipe and synthetic-data generator are promised rather than available.
NVIDIA releases Kumo Tabular for zero-training table prediction
2026-09-29Verified announcementTRIAL
NVIDIA released Kumo Tabular, an open-weight tabular foundation model that predicts classification or regression targets from labeled rows in one forward pass without task-specific training or feature engineering. Three models span 28M to 215M parameters and were pretrained entirely on synthetic tables. NVIDIA reports first-place aggregate results on TabArena, BeyondArena, TALENT and ScoringBench. The weights are ungated under OpenMDW-1.1 and inference code is Apache-2.0, but the training recipe and synthetic-data generator are promised rather than available.
Official · AnnouncementNVIDIA Kumo Tabular release articleOfficial · ReferenceNVIDIA Kumo Tabular model card and weights
Availability: ungated weights plus GPU-native structured-data-models library
What this means for you
For teams maintaining many small gradient-boosted models, Kumo Tabular is a credible candidate for cold-start and low-label baselines. Trial the Medium model on held-out tables and calibration, then compare end-to-end latency and GPU cost with tuned tree ensembles. Do not treat the four leaderboard wins as independent confirmation: they are vendor-run, and performance can degrade outside the training ranges or under context/query distribution shift.
Evidence & related material
Technical details & scores
| Sizes | 28M to 215M parameters; Small, Medium and Large | ||||
|---|---|---|---|---|---|
| Tasks | classification and regression from labeled context rows | ||||
| Training | synthetic Structural Causal Model tables only; about 35M/71M/137M tables | ||||
| Context Scale | training stages extend to 60,000 rows and 100 columns | ||||
| Vendor Benchmarks | ranked first overall on TabArena, BeyondArena, TALENT and ScoringBench | ||||
| Availability | ungated weights plus GPU-native structured-data-models library | ||||
| Licenses |
| ||||
| Limits | numerical/categorical inputs; up to 10 classes per forward pass before error-correcting extension; shift and out-of-range tables can degrade accuracy; training recipe/data generator not yet released |
Technical impact 4
Engineering value 4
Adoption signal 2
Maturity 3
Verification cost 3
Risk 3