NVIDIA's Kumo Relational predicts on raw databases with no training and claims nearly 4 AUROC points over tuned boosted trees, but every number is NVIDIA's own
NVIDIA says the pre-trained model beat supervised baselines on 41 tasks and runs about 430x faster than the top-scoring RelArena entry, at a lower Elo (1,842 vs 1,878). Verdict: cheap to trial on your own tables, too early to retire a boosted-tree pipeline.
Kumo Relational is the first few-shot foundation model to outperform supervised machine learning on common relational benchmarks.
Comprehensive take
A credible, well-documented vendor benchmark claim that is cheap to test: run it on your own multi-table prediction task before changing any production pipeline.
Use cases
Churn, demand and fraud on linked tables
Write a PQL query over users, orders and items instead of hand-building aggregate features; the post links a quickstart notebook.
Replacing per-task XGBoost pipelines
Trial one task side by side with your engineered-feature baseline; NVIDIA reports +7 MRR on SAP SALT out of the box and +13 after fine-tuning.
Agents that forecast outcomes
PQL gives an agent a declarative way to ask a predictive question; the NIM and client library serve it as a service.
Specs
| Context size | Up to 10,000 in-context examples per prediction source |
|---|---|
| Benchmarks | 41 tasks, 15 databases (RelBench V1/V2, SAP SALT, 4DBInfer), none seen in pre-training source |
| RelArena-α | ~17 s per 1,000 samples, Elo 1,842; top entry rt-plurel Elo 1,878 at ~7,300 s source |
| Licence | Code Apache 2.0; weights OpenMDW License Agreement 1.1 source |
| Paper | arXiv 2604.12596, KumoRFM-2, submitted 14 Apr 2026 source |
Insights
- Axion reading: the post says pre-training used 'entirely synthetic' data while the paper abstract says 'synthetic and real-world data'; the descriptions differ and we could not tell which applies to the released weights.
- Axion note: 'about 430x faster' compares against the top-Elo entry, which is more accurate (1,878 vs 1,842); Kumo Relational is best on the accuracy-speed frontier, not best on accuracy.
- Axion note: all results are authors' own; no independent reproduction is cited.
What NVIDIA says it shipped
NVIDIA's Kumo Relational is a pre-trained model that predicts directly on a relational database. You write a predictive query in PQL, for example `PREDICT COUNT(orders.*, 0, 30, days) FOR EACH users.user_id`, the system pulls up to 10,000 historical examples as context, and the model returns predictions in one forward pass with no task-specific training and no flattening of tables into a feature matrix. Code is Apache 2.0, weights are under the OpenMDW License Agreement 1.1, and it runs either through the `structured-data-models` library or as an NVIDIA NIM.
The numbers, and where they come from
All of the following are NVIDIA's own figures from its blog post, drawn from a paper submitted to arXiv on 14 April 2026 (titled KumoRFM-2). Across 41 tasks on 15 databases from RelBench V1/V2, SAP SALT and 4DBInfer, none seen in pre-training, the model is said to beat both supervised relational deep learning models and expert-engineered gradient-boosted tree pipelines. On RelBench V1 classification it is about 4 AUROC points above boosted trees on engineered features; on SAP SALT it is 7 MRR points above XGBoost out of the box and 13 after fine-tuning, and 8% above AutoGluon ensembles in mean reciprocal rank.
The speed claim is the one practitioners should read twice. On the RelArena-α leaderboard the post puts Kumo Relational at roughly 17 seconds per 1,000 samples and an Elo of 1,842, against the top entry, rt-plurel, at Elo 1,878 and about 7,300 seconds. That is the source of the "about 430x faster" line, and it also means Kumo Relational is not the most accurate entry on that board, only the best accuracy-per-second one.
Where the framing needs care
Three things are worth keeping in mind before you rip out an XGBoost pipeline. First, the headline "no training required" is accurate for the model weights but not for the workflow: context generation, a historical replay of database state, and a PQL query still have to be set up, and the post itself recommends starting with a focused query over a few tables. Second, the post says the model is "pre-trained on entirely synthetic relational data", while the abstract of the linked paper says pre-training uses "a large corpus of synthetic and real-world data". The two descriptions differ, and we could not resolve which describes the released weights. Third, the benchmark wins are from the authors; the post links the paper but no independent reproduction exists yet, and NVIDIA concedes that on very large training sets task-specific supervised models may still edge out in-context prediction, which used at most 10,000 examples here.
Who should try it
Teams with churn, demand or fraud targets spread across several linked tables, and with a boosted-tree pipeline that costs weeks of feature engineering per task, get the most from a one-afternoon trial: the post links a quickstart notebook. Teams with one big, stable, heavily tuned tabular model and millions of labelled rows have the least to gain, and should wait for third-party numbers on their kind of data.
Axion take
Treat this as a credible, well-documented benchmark claim from a vendor, not a settled result. The open code, downloadable weights and the 0.2%-of-data and 75%-missing-foreign-key robustness claims make it cheap to test on your own database, which is the only comparison that matters.