Kumo Relational reads a database as-is and predicts without a training run

Most business prediction still runs on hand-built pipelines: an analyst flattens tables into features, trains a model such as XGBoost for one task and maintains it indefinitely. Nvidia’s Kumo Relational, described on the Hugging Face blog, aims to skip that. Users point it at relational tables, state a question in its Predictive Query Language, and it pulls up to 10,000 historical examples as context and answers in a single forward pass. Nvidia reports it beats supervised relational deep learning models and expert-engineered gradient-boosted tree pipelines across 41 tasks on 15 databases it did not see in pre-training, and tops AutoGluon ensembles by 8% in mean reciprocal rank on the SAP SALT enterprise benchmark.

The model ships in Nvidia’s open-source, GPU-native structured-data-models library, with code under Apache 2.0 and weights under the OpenMDW License, and can run as an Nvidia NIM. Nvidia says it scales to databases with more than 500 billion rows and can be fine-tuned when maximum accuracy matters. An earlier hands-on walkthrough of the library found the real risks sit in quiet defaults (zero-training tabular predictions).

Language models replaced the one-model-per-task NLP stack; structured data, the rows and keys behind billing, inventory and risk, has remained largely hand-built. A model that respects tables, foreign keys and timestamps instead of flattening them is a serious approach. Kumo Relational is pre-trained entirely on synthetic relational databases, building on the PluRel approach that members of the Kumo team co-authored, which sidesteps the scarcity of public enterprise schemas. Temporal tasks replay past database states so the model only sees data from before the prediction point, where homemade pipelines often leak. Nvidia also lists limits: very large training sets can still favor task-specific models, and the predictive signal must sit within a few hops of the entity being predicted.

These are benchmarks reported by the team that built the model. Independent replications on data Nvidia did not choose, and production case studies against existing XGBoost pipelines, are the next tests. Nvidia has pursued similar domain-specific approaches elsewhere (Nemotron with Palantir Foundry).

Churn, demand and fraud predictions turn into actions on people: credit holds, flagged accounts, retention offers. Nvidia says the model is built to be called by agents. When an agent can ask a predictive question and act on the answer without a machine-learning team in the loop, speed rises and scrutiny can fall. A bad feature pipeline used to cost weeks of work; a bad query could become an automated decision nobody reviewed, and in-context predictions may be harder to explain afterward. Teams adopting it would be wise to treat each query as a model they own, with validation, leakage checks, drift monitoring and audit. More in our Nvidia coverage, enterprise coverage and AI tools coverage.