NVIDIA releases Kumo models for structured business data
Open models and a library stack for forecasts, anomaly scores, and recommendations on tables. NVIDIA says the weights are on Hugging Face and available as a NIM.
Official source (opens in a new tab)
NVIDIA's 21 October 2026 post introduces the Kumo family, open-weight models and libraries for structured business data. NVIDIA says they help an agent or an application forecast an outcome, score fraud risk, and recommend an action from relational and tabular data, without a separate machine-learning pipeline for every task. Kumo Structured covers relational and tabular prediction, forecasting, and anomaly detection. Kumo Engine is the open library stack for building and deploying those models. NVIDIA says you can start from the pretrained models and fine-tune on your own data when you need to.
NVIDIA says the models are on Hugging Face and packaged as an NVIDIA NIM on build.nvidia.com. The post names early users, including Cognite, Coinbase, Databricks, DoorDash, Revolut, and Tubi, and says Snowflake customers are already using the models. It does not publish parameter counts or benchmark scores.
Who it's for: a team that wants predictions from warehouse tables inside an agent, and can fine-tune if the pretrained model is not enough.
Fits: Best Agent Builder stays n8n. Kumo is a model family for tables, not the workflow tool you hand a team. The verdict is unchanged.