By No Code MBA · Facts checked
- Provider
- NVIDIA
- Announcement date
- Status at last check
- Generally available
- Model family
- Kumo Tabular
- Capabilities
- Text
What changed
Kumo Tabular is an open in-context foundation model for tabular classification and regression: it reads labeled rows as examples and predicts new rows in one forward pass with no training. It comes in Small, Medium and Large sizes spanning 28M to 215M parameters, pretrained on up to 137M synthetic tables with up to 60,000 rows and 100 columns. NVIDIA reports first place on TabArena (Elo 1950), BeyondArena and TALENT.
Access and availability
Weights on Hugging Face at nvidia/Kumo-Tabular under the OpenMDW 1.1 license (commercial use allowed), with code at github.com/NVIDIA/structured-data-models. The library is GPU-native and downloads weights on first use.
Official access documentationWhat you can build
Evaluate it on a held-out tabular task against a strong tree-based baseline, especially when labeled examples are limited.
Limitations
It handles numerical and categorical columns only; text, images and timestamps need preprocessing. NVIDIA warns accuracy may degrade outside training ranges or under distribution shift and advises validating on your own held-out data. Classification covers up to 10 classes per forward pass.