Back to Gap Map Infrastructure / Classic ML & computer vision

TabFM

Google Research
restricted / Overall score: 2.8

TabFM is a tabular foundation model from Google Research that classifies or regresses on a new table without a training step: the user's labeled rows are read as context and predictions for new rows come from a single forward pass. It handles mixed numeric and categorical columns and is pretrained only on synthetic tables. The tabfm package wraps it in scikit-learn-style estimators with JAX and PyTorch backends.

Openness

2 high confidence
2.0
weights
open (1.0.0 and 1.1.0 checkpoints in PyTorch and JAX on the Hub, ungated; the package downloads them on first use)
data
closed (synthetic tables drawn from a structural causal model prior; the generator is not released)
code
partial (inference and estimator code; no pretraining code)
license
Apache-2.0 (code)+TabFM-Non-Commercial-License-v1.0 (weights; non-commercial, non-production use only)

The tabfm code is Apache-2.0, but the pretrained weights the package downloads by default are under Google's TabFM Non-Commercial License, which allows testing, evaluation and research only and bars production use. The synthetic training data cannot be regenerated, since neither the generator nor the pretraining code is released.

Adoption

2 medium confidence
2.0

Hugging Face downloads over the trailing 30 days for the three TabFM checkpoint repositories, most of them for the 1.0.0 PyTorch weights. Monthly PyPI downloads of the tabfm package, which fetches those weights, fall in the same range.

Capability

4 medium confidence
4.0

TabFM reads the user's training rows as context and predicts on new rows without fitting or tuning, which is what TabPFN does. It is one model, limited to ten classes, and does not search over or ensemble other learners.

Verified 2026-11-12