Introducing TabFM: A zero-shot foundation model for tabular data

2026-06-30 · Source: The latest research from Google · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, medium

Summary

Google Research introduced TabFM on June 30, 2026, a new zero-shot foundation model for tabular data classification and regression, now available on Hugging Face and GitHub, and integrated into BigQuery ML. TabFM addresses the traditional bottlenecks of manual feature engineering and hyperparameter optimization in tabular machine learning by reframing prediction as an in-context learning (ICL) problem. Its novel hybrid architecture synthesizes elements from TabPFN and TabICL, employing alternating row and column attention, row compression, and a Transformer for efficient ICL. The model was pre-trained on hundreds of millions of dynamically generated synthetic datasets using structural causal models, overcoming the scarcity of diverse real-world tabular data. Benchmarked on TabArena across 38 classification and 13 regression datasets (700 to 150,000 samples), TabFM and its ensemble variant consistently achieved superior Elo scores compared to heavily tuned, industry-standard supervised algorithms.

Key takeaway

For Data Scientists and ML Engineers building tabular classification or regression models, TabFM significantly streamlines your workflow. You can now achieve high-quality predictions on new datasets without extensive hyperparameter tuning or manual feature engineering. This shifts focus from tedious model preparation to direct analysis, allowing you to deploy robust models faster. Consider integrating TabFM via BigQuery ML or its open-source repos to accelerate your predictive analytics projects.

Key insights

TabFM applies zero-shot in-context learning to tabular data, eliminating manual tuning and feature engineering for classification and regression.

Principles

• Tabular prediction can be reframed as an ICL problem.
• Synthetic data enables large-scale foundation model pre-training.
• Hybrid attention architectures capture complex feature interactions.

Method

TabFM processes entire datasets as a unified prompt, using alternating row/column attention, row compression, then a Transformer for in-context learning on compressed embeddings.

In practice

• Generate high-quality predictions in a single forward pass.
• Access TabFM via Hugging Face, GitHub, or BigQuery ML.
• Utilize TabFM-Ensemble for enhanced performance.

Topics

• Tabular Data
• Foundation Models
• Zero-shot Learning
• In-context Learning
• BigQuery ML
• Synthetic Data Training
• TabArena Benchmark

Code references

• google-research/tabfm

Best for: CTO, VP of Engineering/Data, Director of AI/ML, Data Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• [Webinar Recording] Zero Tuning. State-of-the-Art Predictions. How TabH2O Changes Tabular ML — Tab H2O is a pre-trained, transformer-based tabular foundation model offering high-accuracy predictions without tuning.
• GOTabPFN: From Feature Ordering to Compact Tokenization for Tabular Foundation Models on High-Dimensional Data — GOTabPFN enhances small tabular foundation models for High-Dimensional, Low-Sample Size data via graph-guided feature ordering and neuro-inspired subunit compression.
• Staying Alive: Uncensored Survival Analysis with Tabular Foundation Models — Tabular Foundation Models can perform uncensored survival analysis without dataset-specific training.
• Data Presentation Over Architecture: Resampling Strategies for Credit Risk Prediction with Tabular Foundation Models — Context construction significantly impacts Tabular Foundation Model performance in imbalanced credit risk prediction.
• I Improved My Model… But My Score Got Worse — Cross-validation improvements don't guarantee generalization; simpler, domain-driven features often outperform complex ones.
• Fundamental’s Large Tabular Model NEXUS is now available on Amazon SageMaker JumpStart — NEXUS is a foundation model purpose-built for tabular data, delivering deterministic, scalable predictions without extensive feature engineering.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The latest research from Google.