NVIDIA Kumo: Open-Weight Models for Reasoning Over Structured Enterprise Data

NVIDIA announced the Kumo family of open-weight foundation models at GTC Berlin on October 21, 2026, built to let AI agents reason over structured enterprise data including tabular, relational, time-series and forecast records. The Kumo models are available immediately as open weights on Hugging Face and as NVIDIA NIM microservices on build.nvidia.com, and integrate with the open-source NVIDIA Agent Toolkit alongside Nemotron Retriever and OpenShell enforcement. NVIDIA claims Kumo Tabular delivers up to a 20% improvement in prediction accuracy over standard gradient-boosted decision trees for agentic workloads, a company-reported figure.

On October 21, 2026, at GTC Berlin https://blogs.nvidia.com/blog/gtc-berlin-2026-keynote/ , NVIDIA announced the Kumo family of foundation models, designed specifically to connect AI agents to structured enterprise data. Part of the broader Agent Knowledge Foundation https://developer.nvidia.com/topics/ai/agent-knowledge-foundation , Kumo addresses the ‘Data Plumbing’ challenge identified in the TGVP infrastructure report https://forkast.news/tgvp-infrastructure-not-model-quality-is-the-primary-bottleneck-to-enterprise-ai-agent-adoption/ , where the inability to reason over databases remains a primary barrier to autonomous agent deployment. The Kumo family includes specialized models for Tabular , Relational , Time Series , and Forecast data. While the company’s Nemotron series handles unstructured documents, Kumo provides the prediction layer over structured records. In the announced workflow, the Nemotron tools parse documents, embed queries and rerank information, while the new Kumo models make predictions from structured data – classification and regression, in the model card’s terms. As CEO Jensen Huang framed it at the keynote: “Ultimately, what turns these super agents, these agents into super agents, is that you give them access to your company’s knowledge.” Availability is immediate: Kumo models are available as open weights https://forkast.news/glossary/open-weights-models/ on Hugging Face https://huggingface.co/nvidia/Kumo-Tabular and packaged as NVIDIA NIM microservices on build.nvidia.com for production inference. The integration into the open-source NVIDIA Agent Toolkit allows developers to combine Kumo with Nemotron Retriever and OpenShell enforcement https://forkast.news/nvidia-pushes-agent-enforcement-into-the-silicon-and-publishes-the-blueprint/ , creating a full-stack governance and knowledge foundation. NVIDIA claims the Kumo Tabular model provides up to a 20% improvement in prediction accuracy over standard gradient-boosted decision trees for agentic workloads company-reported . For enterprise architects, Kumo represents the verticalization of the data layer. By building a dedicated moat around how agents access structured data, NVIDIA is positioning itself as the essential infrastructure provider for both the safety OpenShell and the knowledge Kumo of the agentic economy. This release suggests that the future of agentic RAG https://forkast.news/glossary/rag-retrieval-augmented-generation/ is moving beyond simple document retrieval and toward a unified framework that treats structured databases as first-class citizens in the agent’s context.