Google's TimesFM-3 learns to see the future in a single forward pass

The new generation model can account for future events

Google Research has unveiled the third generation of its time-series foundation model, now scaling to 330 million parameters. While older versions could only extrapolate a single variable from its own history, TimesFM-3 can observe several correlating graphs at once. Crucially, the model accepts known future events (like scheduled marketing or weather forecasts) and bakes them into the output before they actually happen.

Contiguous Patch Masking resolves the entire horizon at once

From an infrastructure perspective, the shift is in how the model outputs results. Instead of autoregressive decoding, where each subsequent value is calculated from the previous one, Google introduces Contiguous Patch Masking. The model sets up empty windows for the entire future timeline and fills them in a single forward pass using full variate attention.

Native integration with enterprise data locks companies into the cloud

While the model promises salvation from task-specific fine-tuning, the reality of cloud deployment means platform lock-in. The critical integration test arrives in the coming weeks with its BigQuery rollout, where the feature will hit enterprise data via a simple AI.FORECAST SQL command.

Handing inventory over to an algorithm will be the real test

The future signal of success lies not in benchmarks, where the model currently leads, but in corporate adoption. The proof will be whether retail chains dare to hand over their inventory management to a non-linear model instead of sticking to three-year historical averages in Excel.

Lilith's verdict

Ditching Excel and letting an algorithm fill fifty calendar cells forward in one swoop sounds appealing. But the day the model miscalculates a holiday and all of Europe runs out of bread, the blame falls squarely on the human who ran AI.FORECAST.