Share
What is NVIDIA Kumo?
NVIDIA Kumo (formerly Kumo.ai) brings together the relational prediction technology Kumo built, now part of NVIDIA. Its core is Kumo Relational, a pre-trained foundation model that predicts from several linked tables, so you do not have to train and maintain a separate model for every question.
How it works
For each request, the model pulls in table relationships and relevant historical rows as context. You state the target in Predictive Query Language (PQL), a light SQL-style syntax, and get the answer back as a pandas DataFrame through the kumo-relational-client Python SDK. Data can come from Snowflake, Databricks, DuckDB, SQLite or Amazon S3.
What it predicts
- Classification and regression: categories, yes/no outcomes and numeric values, such as churn risk.
- Forecasting of numeric values over a future horizon.
- Link prediction to rank or recommend related entities.
Predictions come with explanations, and batch prediction is supported.
Deployment
Kumo Relational runs as an NVIDIA NIM microservice: deploy the container on your own NVIDIA GPUs or call a hosted endpoint from the NVIDIA API Catalog. It is aimed at data and ML teams working with customer records, transactions or product catalogs.
Pricing
NVIDIA does not list a price for Kumo Relational; terms depend on how you deploy it and are arranged with the company.
NVIDIA Kumo features and use cases
How to use NVIDIA Kumo
NVIDIA Kumo FAQ
What happened to Kumo.ai?
Its technology moved to NVIDIA and is now presented as NVIDIA Kumo, with Kumo Relational as the main model and docs hosted by NVIDIA.
What is a relational foundation model?
A pre-trained model that understands data spread across linked tables and can answer different prediction questions without task-specific training.
Do I need to train a model for each prediction?
No. That is the point: each request uses the relationships and history in your tables as context.
What is PQL?
Predictive Query Language, a light SQL-like syntax for naming the value you want to predict and the entities to score.
Which prediction types are supported?
Classification and regression, time-based forecasting and link prediction for recommendations or rankings.
Where does it run?
As an NVIDIA NIM container on your own NVIDIA GPUs, or through a hosted endpoint in the NVIDIA API Catalog.
Which data sources can it read?
The docs cover Snowflake, Databricks, DuckDB, SQLite and Amazon S3.
How much does it cost?
NVIDIA does not publish a price; terms depend on the deployment you choose.
NVIDIA Kumo pricing and plans
Price not published
- NVIDIA does not list rates for Kumo Relational.
- It ships as a self-hosted NIM microservice on NVIDIA GPUs or as a hosted endpoint in the NVIDIA API Catalog.
- Usage and licensing terms are arranged with NVIDIA.
NVIDIA Kumo Screenshots
NVIDIA Kumo user reviews
Write a review
Checking your session...
NVIDIA Kumo Reviews
Loading reviews...
NVIDIA Kumo alternatives
Mel By Skymel
Tracks open-web signals to build knowledge graphs. Generates intelligence reports and databases, and monitors relevant changes for any target.
Velivo One
Monitor AI engine mentions, verify accuracy, and manage your brand's SEO, local presence, and social media to optimize visibility and digital reputation.
ParseCrate
Open-source, self-hosted document processing API. Converts PDFs and Office files into structured content with OCR and table detection for AI applications.