- How tabular foundation models work: column embeddings, target-aware attention, and in-context learning in a single forward pass
- Why a pre-trained, zero-tuning model can rival XGBoost and GBM without any feature engineering
- Where TabH2O holds accuracy at scale, up to hundreds of thousands of rows and 1,700+ columns
- The tasks it handles: supervised regression and classification, plus anomaly detection and clustering
- How to run it today, from the hosted API and website to an Excel or Google Sheets plugin and an agent skill file
- What enterprise deployment looks like, including on-prem and virtual private cloud
Overview
TabH2O is H2O.ai's tabular foundation model: a transformer pre-trained on 6 million synthetic datasets that predicts on your data with zero tuning, no feature engineering, and no training run. When you send it a table, it runs column embeddings, row-and-column attention, and in-context learning in a single pass, learning from your data at inference time much the way a language model reads a prompt. At roughly 29 million parameters it stays small and fast to deploy, yet holds best-in-class accuracy against gradient-boosted trees like XGBoost and GBM, scaling to hundreds of thousands of rows and well over a thousand columns. In this session, Kaggle Grandmaster Mark Landry walks through how the model works, shows live benchmarks, and goes from a raw spreadsheet to predictions in minutes.