NVIDIA Releases Kumo Tabular: Open Tabular Foundation Models That Predict New Rows in a Single Forward Pass
NVIDIA open-sources tabular foundation models that skip training entirely—context in, predictions out in one pass.
Why it matters
NVIDIA's Kumo Tabular extends the in-context learning paradigm (TabPFN, TabICL) to tabular data at scale. No training, no hyperparameter tuning, no feature engineering required. This matters for practitioners deploying structured-data ML at speed, but the capability claim and real-world accuracy vs. traditional methods remain unvalidated in this source.
The key facts
13 to knowNVIDIA releases Kumo Tabular family of tabular foundation models
In-context learning approach: labeled rows as context, single forward pass prediction
No training, hyperparameter tuning, or feature engineering required
Follows TabPFN and TabICL precedent
Open-weight models (implied by 'open' in title)
Classification and regression capability
Publication date: 30 September 2026 (via MarkTechPost)
No independent validation data, benchmarks, or accuracy comparisons provided in article
Model family: Kumo Tabular (open release, weights available)
Capability: in-context learning on tabular data — takes labeled rows as context, predicts new rows in single forward pass
Lineage: similar architecture to TabPFN and TabICL (established in-context learning for tabular tasks)
Release status: open model (weights made available, not just API)
Source: MarkTechPost summary of NVIDIA announcement; original NVIDIA release not independently verified in brief
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: NVIDIA has released Kumo Tabular, a new family of tabular foundation models (TFMs) for classification and regression. If you have followed TabPFN or TabICL, the setup will look familiar. The model takes labeled rows as context and predicts new rows in one forward pass. There is no training, no…