NVIDIA Kumo Relational predicts on databases untrained

What changed?

On October 21 NVIDIA introduced Kumo Relational (Hugging Face blog), a pre-trained relational foundation model. You connect a database, describe what to predict and get results in a single forward pass. There is no per-task training, no feature engineering and no flattening of tables into one row.

We also covered the Kumo Forecast time-series model today. Relational is a separate member of the family that works over tables, foreign keys and timestamps.

How it works

• In-context learning: for each prediction it gathers up to 10,000 historical examples as context, each a subgraph around an entity at a point in time paired with its real outcome.
• Predictive Query Language (PQL): a declarative language where you state what to predict, not how. For temporal tasks the model replays past database states, which prevents data leakage.
• Two stages: a light network uses column and row attention within each table to filter noise, then a larger one links tables with graph attention over key relationships.
• Pre-training: entirely on synthetic relational databases built on the PluRel generator. None of the benchmark datasets were seen in pre-training.

What the numbers show

NVIDIA calls it the first few-shot foundation model to beat supervised machine learning on common relational benchmarks. It was tested on 41 tasks across 15 databases from RelBench V1 and V2, SAP SALT and 4DBInfer.

• On RelBench V1 classification it beats gradient-boosted trees on expert-engineered features by nearly 4 AUROC points on average.
• On SAP SALT it beats XGBoost by 7 MRR points out of the box and by 13 after fine-tuning. It leads AutoGluon ensembles by 8%, or 16% after fine-tuning.
• It scores about 25% higher than recent tabular foundation models on SAP SALT.
• On the RelArena-α leaderboard it sits on the Pareto front at Elo 1,842 and about 17 seconds per 1,000 samples, roughly 430 times faster than the top entry, rt-plurel (Elo 1,878).

Source: NVIDIA, Hugging Face blog (RelArena-α leaderboard, October 21, 2026)

How to try it

The model runs in NVIDIA's open, GPU-native structured-data-models Python library (pip install structured-data-models), which downloads the weights from Hugging Face on first use. It can also be deployed as an NVIDIA NIM, with structured-data-models-client providing a PQL-based client. A quickstart notebook and a Databricks example are available.

Code is released under Apache 2.0 and weights under the OpenMDW License 1.1. NVIDIA says it scales to databases with more than 500 billion rows.

Limits to know

It works best on data with meaningful relationships and temporal dynamics, such as user behavior over time. With very large training sets a task-specific supervised model may still edge it out, which is what fine-tuning is for. Predictive signals must be reachable within a few hops, so start with a focused query over a handful of relevant tables.

If you run XGBoost pipelines for churn or demand, benchmark both on a held-out period. If Kumo Relational keeps up, the next prediction really becomes a query instead of a project.