Google's TabFM learns to predict from data it has never trained on

Key points

• Google Research published TabFM in 2025, a foundation model that makes predictions on brand-new datasets without any per-dataset training.
• TabFM matched or beat heavily tuned traditional models on 51 public benchmark datasets across 38 classification and 13 regression tasks.
• The model was trained entirely on hundreds of millions of synthetic, computer-generated datasets, so it never ingested real customer or business data.
• Inference is computationally heavier than traditional approaches, which teams must plan for.
• Google researcher Weihao Kong told VentureBeat the model's chief value is velocity: high-quality predictions without a dedicated data science team.

Most business data lives in spreadsheets and financial ledgers. Getting an AI model to make useful predictions from that kind of data, whether forecasting customer cancellations or spotting fraudulent transactions, has always required a lot of manual work first.

A team at Google Research has now published an approach called TabFM, short for Tabular Foundation Model, a general-purpose AI system built to handle the row-and-column data you'd find in any spreadsheet.

TabFM can predict outcomes on a dataset it has never seen before, without any custom training. You hand it your historical records and the new rows you want to understand, and it returns a prediction in one step.

Why is this harder than it sounds?

Traditional machine learning requires a specialist to build a bespoke pipeline for every new dataset: cleaning messy data, encoding categories as numbers, running dozens of test configurations, then monitoring the model after launch because real-world data keeps shifting. It's slow, expensive, and demands ongoing attention.

Language models, the technology behind chatbots like ChatGPT and Claude, are no shortcut here. They were built for text, not grids. A medium-sized table with a few thousand rows quickly fills their memory. Numbers get mangled because the model chops them into odd fragments. Flattening a two-dimensional grid into a single line of text causes the model to lose track of which value belongs to which column.

"That's why, today, it is far more effective to use an LLM to write the code that handles feature engineering than to ask the LLM to read the table itself," Kong told VentureBeat.

How does TabFM actually work?

TabFM keeps the table's grid structure intact throughout. It alternates attention, a technique where the model checks how each value relates to its neighbours, across rows then columns. It then compresses each row into a compact summary before making its final prediction.

The design borrows from two earlier models: TabPFN, developed by Prior Labs, which proved the concept but struggled with large tables, and TabICL, from France's national digital research institute, which added row compression to handle bigger datasets.

TabFM never trained on real business data. Google built it entirely on hundreds of millions of computer-generated synthetic datasets, so no private customer records were used in training.

How well does it perform?

On TabArena, a standard evaluation suite covering 51 diverse real-world datasets, TabFM's zero-shot predictions, meaning predictions made with no prior exposure to that data, matched or outperformed heavily tuned traditional models. Google is careful to note this does not mean TabFM wins every head-to-head against a specialist model built for one specific task.

For teams chasing maximum accuracy, Google also introduced a "TabFM-Ensemble" configuration that runs the model through 32 distinct variations and blends the results.

What's the real trade-off?

Traditional models are slow to build but fast to run. TabFM flips that. There's no training phase, but every prediction requires the model to re-read all your historical records as context. That's computationally heavy. For applications needing single-digit millisecond responses, like a real-time fraud check at a payment terminal, the latency may be a problem today.

For a small team that needs a working baseline by next week rather than next quarter, the calculus looks very different. That tension between inference cost and deployment speed is something we've tracked across recent AI cost stories: cheap to start is not always cheap to run.

The honest read on TabFM: the zero-shot benchmark results are genuinely impressive, and the synthetic-only training approach is a clean answer to data-privacy concerns. The inference cost is real and not a footnote. Watch whether Google releases a latency profile against production-grade fraud or churn workloads before enterprise teams commit.