Google's TabFM handles any table without retraining

Google released TabFM, a foundation model for tabular data. Unlike older approaches that retrain for every dataset, it works zero-shot on any table. It handles everything from customer records to sensor readings.

The model treats each table as a sequence — rows and columns become tokens the model can reason over. You feed it a table, it spits out predictions. No fine-tuning required.

This is the kind of thing the cousin running a warehouse management side business would use. Tables of inventory, orders, supplier lead times — now one model covers them all.

Why this matters for us: the gente who run small operations with spreadsheets and databases can finally get ML without hiring data scientists — the model learns once and just works.