Google brings TabFM predictions to BigQuery in preview

Google has introduced the TabFM model in BigQuery. It is available in preview.

The launch adds a pre-trained regression and classification model for tabular data to BigQuery, Google's cloud data warehouse. It is designed to let users generate predictions through SQL instead of building and deploying a separate custom model.

Predictive analytics tasks such as churn forecasting, fraud scoring and purchase intent modelling have often required businesses to train models with tools such as XGBoost, random forests or deep neural networks. That process can involve feature engineering, hyperparameter tuning, deployment and retraining, which tends to favour teams with specialist data science resources.

TabFM takes a different approach. The model uses in-context learning, meaning users provide labelled historical data and a new table for prediction, and the system returns results through a single SQL statement.

In BigQuery, the model is exposed through two built-in SQL functions: AI.PREDICT and AI.EVALUATE. The first produces predictions from historical and new data, while the second assesses performance against a test set using standard metrics for classification and regression tasks.

The model automatically determines whether a task is classification or regression based on the target label's data type. It also handles routine data preparation tasks, including missing values and categorical encoding, inside BigQuery.

Benchmark claims

Google said TabFM was developed by Google Research and described it as a foundation model for tabular data. It said the model outperformed both established machine learning approaches and other tabular foundation models in evaluations on the TabArena benchmark.

That claim sits within a broader push by large technology groups to apply foundation model techniques beyond text, images and code. Tabular data, which underpins many business applications from customer records to transaction logs, remains one of the most common data types in corporate analytics.

The model is aimed at users who want prediction tools inside existing analytics workflows rather than through a separate machine learning stack. By placing the feature in BigQuery, Google is targeting analysts, developers and data teams already working in SQL.

How it fits

Google positioned TabFM as a complement to existing options rather than a replacement for traditional machine learning models. It said the model is intended for cases where teams want fast predictive output without extensive model development, especially when historical datasets are small to medium-sized or change frequently.

For larger historical datasets, or when users want detailed control over tuning and feature-level explainability, Google said traditional approaches such as XGBoost may still be a better fit. That suggests the company expects both methods to coexist in BigQuery depending on the use case.

The release also reflects broader competition in cloud data platforms, where providers are adding more built-in artificial intelligence features to keep users inside a single environment for storage, analysis and modelling. In practical terms, integrated SQL-based prediction tools can reduce the need to move data between warehouse products and external machine learning systems.

Operational focus

Google said BigQuery uses distributed, parallelised inference to manage the computational demands of tabular foundation models. According to the company, that allows the system to process inference tables containing up to millions of rows in minutes.

It also said BigQuery uses training-data sampling and distributed execution to improve performance and manage resource use. Those implementation details matter because tabular machine learning workloads can become expensive and slow when datasets are large or frequently refreshed.

The addition of predictive functions in SQL may also widen access to machine learning-style analysis across business teams. Rather than relying on data scientists to build, test and deploy a model, users can work from historical labelled datasets within a familiar query environment.

At the same time, the preview status indicates the feature is still at an early commercial stage. Users in regulated or highly sensitive settings may still weigh questions around accuracy, explainability and operational controls before replacing established modelling methods.

Google said TabFM can also be used alongside the BigQuery MCP server for agent-based business applications, extending its use beyond standard dashboard and reporting environments. The central pitch, however, remains simpler predictive analysis on tabular data through SQL, without a separate model training cycle.