TabFM

Google's tabular foundation model: zero-shot prediction on unseen tables via in-context learning, trained "entirely on hundreds of millions of synthetic datasets" generated from structural causal models — no real-world data. Alternating row/column-attention encoder feeding a 24-block causal ICL transformer with 400M parameters (per the arXiv technical report of 2026-09-29). Beats tuned gradient-boosted trees on TabArena; the report puts zero-shot TabFM first among default tabular foundation models across all 51 TabArena datasets and ahead of tuned AutoML pipelines, with two extensions on the same frozen weights (TabFM+ ensembling and calibration, LLM-guided TabFM-Auto). Weights under the TabFM Non-Commercial License v1.0; code Apache-2.0. Extends the TabPFN line of prior-fitted tabular models to Google scale.