TabH2O
A tabular foundation model that predicts without training. Send labeled and unlabeled rows, and get classification, regression, and forecasts back in seconds.
- TabH2O is a foundation model for tabular prediction: send labeled training rows and unlabeled rows, get predictions back with no training run.
- Under three seconds on 15,000 rows; up to 500,000 rows on one GPU; classification, regression, and time-series forecasting.
- Free tier: 2 requests per minute, 20 per day, 500 per month, 100,000 rows, 100 columns, no credit card. Data is never stored.
- TabH2O launched at Dell Technologies World in May 2026 and is validated in the Dell AI Factory with NVIDIA; it also runs on AMD.
- Weights are not released; access is the hosted API, an on-prem service, or inside H2O Cascade for air-gapped use.
What it does, and what it skips.
Classical ML trains a model per dataset. TabH2O is one model that reads your labeled rows as context and predicts the rest. The steps it removes are the steps projects stall on.
- Classification and regressionBinary, multiclass, and regression on tabular data. Send rows, get predictions back with calibrated probabilities.
- Time-series forecastingForecast straight from historical rows, with no training pipeline to build, schedule, or maintain.
- Clustering and anomaly detectionThe same in-context approach applied to unsupervised tasks, including imputation. Available on the paid plans.
- No training, no tuningNo hyperparameters, no feature engineering, no model artifact to version or deploy. Weeks of work become seconds.
- Excel and SheetsAdd-ins for Excel and Google Sheets, so analysts get predictions right where they already work.
- Agent-callableA published SKILL.md lets the H2O AI Super Agent call TabH2O as a tool, fed by the Feature Store online store.
Rows in, predictions out.
- 01Prepare two tablesLabeled training rows and the unlabeled rows to predict. Up to 100 columns on the free tier.
- 02Send themREST API, Python, the Excel or Sheets add-in, or a call from an agent.
- 03Get predictionsSeconds later: probabilities, forecasts, or cluster labels, ready to use.
- 04Nothing to deployNo artifact and no serving stack, and data is never stored by the hosted service.
A parallel inference track beside MLOps.
TabH2O is not a training engine. The Feature Store feeds it directly and predictions come out, with no model artifact and nothing for MLOps to deploy. Highlighted is TabH2O.
What ships today
The prediction step an agent can take without waiting for a model.
When the H2O AI Super Agent needs a forecast or a score inside a larger plan, TabH2O answers in seconds, with no experiment to schedule and no model to deploy. When the case needs a fully trained, documented model, the agent uses H2O Driverless AI instead.
You operate it, or H2O.ai operates it for you.
Same platform either way: the same capabilities, the same governance, and the same audit trail. The choice stays open, so you can switch between them as the business changes, without rebuilding what runs on top.
H2O AI Cloud for Government
H2O AI Cloud for Government is FedRAMP High certified, delivering the highest standard of cloud security so federal agencies can harness predictive and generative AI without compromising data sovereignty or compliance.
Learn TabH2O
:quality(75))
What's new on TabH2O
Questions people ask before they call us.
TabH2O is H2O.ai's tabular foundation model. It is pretrained on millions of synthetic datasets and predicts on your table by reading labeled rows as context, with no training, tuning, or feature engineering. It returns predictions in under three seconds on 15,000 rows.
Try it on your own table today.
Free tier, no training. Or bring a forecasting problem and we will benchmark TabH2O against your current model, on your data.