Kumo Relational

Kumo Relational

Predicts classes or numbers for table rows by learning from example rows and the tables linked to them. It suits data scientists who work with relational data.

Made by nvidia on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Kumo Relational is a pretrained model from NVIDIA for classification and regression on tables that link to each other. It learns from the examples you pass in, with no training run of your own. It builds on KumoRFM-2 and mixes row embeddings with message passing across related tables.

Who it suits: It suits data scientists and engineers who predict outcomes from linked tables, such as a main table plus tables that refer to its rows.

Industry | Who uses it | How they use it | Example
Data science | Analysts with relational data | Predict a label for new rows from labeled example rows | Estimating a class for each new record in a main table
Software | Engineers on data products | Add table predictions to an app through a Python package | Calling the model from a service with context and query tables
Research | Machine learning researchers | Study in-context learning on related tables | Comparing estimator counts on one dataset

People whose data is text, images or a single flat table can look at other models first.

Where to use it:

• Python on Windows, macOS or Linux
• A computer with a CUDA GPU, as in the card's sample
• The structured-data-models package, which runs inference

Use case | Who | What they get back
Classification | Data scientists | Class probabilities for your query rows
Regression | Analysts | Predicted numbers for your query rows

How to set it up:

• Python on your computer
• A CUDA-capable GPU, since the sample sets the device to cuda
• Context tables with known labels, query tables, and any related tables

1. Install the Hugging Face hub package.
2. Download the model files.
3. Install the structured-data-models package.

pip install -U huggingface_hub

hf download nvidia/Kumo-Relational

pip install structured-data-models

How to use it: Load the model for a task, then call it with your labeled context rows, your query rows and their related tables. For classification it returns class probabilities.

import sdm

model = sdm.models.KumoRelational(task="classification", device="cuda")
probs = model(
    x_context=x_context,
    y_context=y_context,
    x_query=x_query,
    related_context_tables=related_context,
    related_query_tables=related_query,
    num_estimators=8,
)

Tool | What it does
sdm.models.KumoRelational | Loads the pretrained model for a classification or regression task
x_context and y_context | Your example rows and their known labels
x_query | The rows you want predictions for
related tables | Linked tables for the context rows and the query rows

Example: You hold a table of customers with known outcomes, plus linked tables for their activity. You pass those as context, add the new customers as query rows, and ask for a classification. The call returns class probabilities for each new customer.

For newcomers and experts: In-context learning means the model reads your labeled examples at prediction time instead of being retrained. A good first try is a small labeled table with one linked table, run through the sample call.

Underneath, it starts from pretrained TabICLv2 weights that were finetuned. The design mixes embeddings of rows with message passing across related tables. The num_estimators setting in the sample is 8. The card says the weights are released under OpenMDW 1.1.

Labeled context rows and linked tables
Query rows and their linked tables
Model reads context in-context
Class probabilities come back

Good to know:

• License: OpenMDW 1.1 for the weights. The model finetunes TabICLv2 weights under BSD-3-Clause, with notices kept in a THIRD-PARTY-NOTICES.txt file.
• Cost to run: none named. The steps need no API key or paid service.
• Data it reaches: it downloads files from the Hugging Face Hub, then works on the tables you pass in.
• Not for: text or image inputs, which the card does not mention.

On Hugging Face, 13 November 2026

• Downloads in the last 30 days: 11,526
• Likes: 47
• License: openmdw-1.1
• Based on: jingang/TabICL
• Tags: kumo_relational, relational-foundation-model, structured-data-models

Links

• Model: https://huggingface.co/nvidia/Kumo-Relational
• Files: https://huggingface.co/nvidia/Kumo-Relational/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download nvidia/Kumo-Relational

Then follow the usage in its model card. Nothing in this kit runs by itself.