NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required

Most business prediction still runs on hand-built pipelines: an analyst flattens tables into features, trains a model like XGBoost for one task, and maintains it forever. NVIDIA's Kumo Relational says it can skip that. Point it at relational tables, state the 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 none of which it saw in pre-training, and tops AutoGluon ensembles by 8% in mean reciprocal rank on the SAP SALT enterprise benchmark. If that holds outside the lab, a new prediction stops being a project and becomes a query.

We think this is one of the more consequential quiet releases of the week. Language models made the one-model-per-task NLP stack obsolete. Structured data, the rows and keys that actually run billing, inventory and risk, has stayed stubbornly artisanal. A model that respects tables, foreign keys and timestamps instead of flattening them is the right bet.

The design choices read as serious. It is pre-trained entirely on synthetic relational databases, which sidesteps the fact that real enterprise schemas are rarely public. Temporal tasks replay past database states so the model only sees data from before the prediction point, which is exactly where homemade pipelines tend to leak. And NVIDIA publishes its own limits: very large training sets can still favor task-specific models, and the signal has to sit within a few hops of the entity being predicted.

Still, I'd hold the champagne. These are vendor benchmarks reported by the team that built the model, and leaderboards reward what they measure. The real test is a messy production warehouse with odd schemas, drifting definitions and business owners who need to explain a decision.

That last point is our main worry. Churn, demand and fraud predictions turn into actions on real 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 with no ML team in the loop, speed goes up and scrutiny goes down. The cost of a bad feature pipeline used to be weeks of work. The cost of a bad query could be a quiet, automated decision nobody reviewed.

Our read: useful, plausibly a real step, and the governance has to move with it. Teams adopting it should treat every query as a model they own, with validation and audit, not a lookup.

Kumo Relational 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 also 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. Its synthetic training data builds on the PluRel approach, which members of the Kumo team co-authored.