Overview
Kumo Tabular is a structured-data foundation model for predictive AI on a single table. It predicts in context: give the model labelled rows and the rows you want scored, and it answers in one pass. There is no training loop, no hyperparameter search, and no per-dataset artifact to manage.
You access Kumo Tabular through structured-data-models-client (pip install structured-data-models-client, imported as sdm_client), NVIDIA’s open-source Python client for structured-data foundation model NIMs. The same client serves both Kumo Tabular and Kumo Relational — the model you want is an argument to predict(), not a different package.
The current Kumo Tabular NIM uses the Kumo Tabular large model. To use the small or medium model, you can package the model (weights found on HuggingFace) and deploy.
The one idea
The model is an argument, not a method name.
One call answers with Kumo Tabular. A request built for another model is refused before anything is sent, because what a model may be asked is declared as data rather than discovered from a failed call.
What you can predict
When not to use it
- You want to train a model. Kumo Tabular does not train. If you need a model fitted to your data, this is the wrong tool.
- Your tables are related. Kumo Tabular reads one table and no relationships. For several related tables joined as a graph, see Kumo Relational.
- Your data cannot leave your environment. It can, but only to a NIM you run yourself. Run one and point the client at it.
- You want inference in-process. Everything here is an HTTP call to a NIM.