[DigitalToday reporter Chi-gyu Hwang] Nvidia has recently released its open foundation model for tabular data prediction, Kumo Tabular, on the open-source AI model-sharing platform Hugging Face.

Kumo Tabular reads tables that include answers and directly predicts values for rows with blank answers. It requires no separate training, tuning or feature engineering and supports both classification and regression.

The Kumo Tabular model is offered in 3 versions with 28 million to 215 million parameters. It uses the OpenMDW-1.1 licence, allowing commercial use. Nvidia said it ranked first across 4 benchmarks: TabArena, BeyondArena, Talent and ScoringBench.

Companies have used gradient-boosted trees models over the past 20 years to predict customer churn, demand and pricing. This approach requires label collection, feature design and validation to be redone from scratch whenever the question changes.

Kumo Tabular applies the in-context learning used by large language models to tables, with the model reading a table containing recorded outcomes as context and directly predicting values for new rows. Only synthetic tables were used for training. Nvidia explained it created random causal graphs with a structural causal model and generated tables from them.

Nvidia also plans to soon release its training method and the synthetic data generator.

Kumo Tabular appears to be based on the technology of enterprise predictive AI software startup Kumo AI, which Nvidia acquired in June.

Kumo AI has developed AI models that answer predictive questions based on companies' structured data such as customer information and payment data. It specialises in structured data analysis that is difficult for general large language models to handle. Kumo launched its latest model, KumoRFM-2, in April, and counts DoorDash, Reddit, Databricks and Snowflake among its customers and partners.