Google is adding a pre-trained foundation model for tabular data to BigQuery, allowing enterprise teams to generate predictions from structured data without having to train and deploy a machine learning model from scratch.
The TabFM model, which has been developed by Google Research and is now available in preview, can perform classification and regression, two common types of predictive machine learning.
While classification is used to predict a category or outcome, for example, whether a customer is likely to churn or whether a transaction may be fraudulent, regression is typically used to predict a numerical value, such as the expected value of a customer or the amount of a future claim. Traditionally, enterprise teams in a typical machine learning workflow would need to train and deploy a separate ML model for each use case they want to pursue. In contrast, TabFM uses historical data as in-context examples to make predictions, the company wrote in a blog post.
That means teams can provide labeled historical data, where the outcome is already known, along with the new data they want to make predictions about, and use BigQuery’s AI.PREDICT function to generate those predictions.
The function uses the historical data as examples to identify patterns and then applies what it has learned to the new data, without requiring teams to separately train and deploy a machine learning model.
TabFM, according to Google, can also automatically work with data expressed as categories, such as customer type or product category, as well as datasets with missing values, reducing the preparation required before making predictions.
Further, the company is providing an AI.EVALUATE function to assess the model’s performance on the prediction task. Teams can use it to compare TabFM’s predictions with the known outcomes in their data and measure how accurately the model performed, Google wrote, adding that this gives teams a way to evaluate whether the predictions are useful for a particular use case.
For enterprises, the addition of TabFM could make predictive analytics more accessible to teams that do not have extensive ML expertise, while reducing some of the engineering and infrastructure work involved in developing and maintaining conventional models, said Amit Kumar Jena, AI development head at IT consulting firm Kanerika. “Someone who knows SQL can now generate a classification or regression prediction directly against tables already in BigQuery, using the two new functions, without training, tuning, or deploying a model first, and without moving data out to a separate ML platform, which also simplifies governance and cuts the infrastructure cost of running a parallel platform,” Jena explained.
That, according to Manoj Chandra Jha, principal analyst at Nord-IQ Research, translates to TabFM collapsing the traditional multi-persona workflow, stretching from an analyst, a data scientist, and a ML engineer, into a single SQL call, thereby potentially reducing headcount requirements.
However, the new model has limitations that enterprises should be aware of before replacing conventional ML workflows.
Google’s own guidance and documentation still point to XGBoost, which is a widely used machine-learning algorithm for making predictions from structured or tabular data, and custom models for very large datasets, datasets with more features than TabFM’s current limit of 20 feature columns.
“Use traditional models like XGBoost when you have very large historical datasets, require complete control over custom hyperparameter tuning, have a high number of features that exceed current limits of TabFM, or need feature-importance explainability, i.e., which of the input features contributed most to the prediction,” the company wrote.
More so because, in large datasets and high-frequency workloads, the cost advantage can diminish, analysts pointed out.
While TabFM can reduce the cost of setting up a model-training pipeline, Google’s plans to move TabFM to token-based pricing on top of standard BigQuery charges starting October 30, 2026, could result in higher costs for workloads that generate predictions at scale compared with a trained and cached model, Jena said. Enterprises should therefore pilot TabFM on their own workloads, run the numbers based on their actual query volumes, and compare the costs with their existing ML infrastructure before shifting significant parts of their ML budgets to the service, Jha advised.
In other words, that makes TabFM more of a workload-specific tool than a wholesale replacement for traditional ML models, Jha said, adding that it is better suited to workloads such as ad hoc analysis, experimentation and lower-frequency business use cases.