Hand the older model a promotions calendar and nothing happens. Hand TimesFM-3 the same calendar and it starts penciling in roughly 20 percent more units sold on every promotion day.
That difference is the entire case for the forecasting model Google Research just released, and it is a narrower case than the launch language implies.
Through TimesFM-2.5, which arrived in September 2025, every member of this family was confined to one data series at a time. Hand it daily ice cream sales and daily ice cream sales came back, projected forward. The heat wave, the waffle cone figures and the Friday promotion already on the books were all invisible to it.
The change is in how many things it is permitted to consider
The example Google reaches for is a retail chain trying to forecast ice cream sales. A forecast that is actually useful there does not rest on a single number. It draws on neighboring products such as waffle cones or syrup, along with earlier foot traffic, weather, discount campaigns and holidays.
Three categories of supplementary data now feed the model. It forecasts several related variables simultaneously, different flavors being the obvious case. It ingests factors that are known only for the past, historical foot traffic among them. And it accepts known future events, meaning scheduled discounts or weather forecasts.
Multivariate support is the actual headline here. The rest of the architecture exists to serve it.
Two passes, pointed in different directions
A Transformer still sits underneath, the same base its predecessors used. Thirty-two consecutive data points are bundled into one patch, and each series is normalized onto a shared scale, which lets measurements of wildly different magnitudes sit side by side without one swamping the other.
From there it switches directions. Along the time axis it looks for patterns within a single series, using past values only, so information from the future cannot leak backward. Across series it lines up every variable at a given moment and learns how they relate, which is the mechanism that picks up a discount on one product cannibalizing or lifting another.
The model carries 330 million parameters and was trained, Google says, on real and synthetic time series amounting to more than one trillion data points. As with the earlier releases, it operates zero-shot, requiring no additional training for a new task.
The slow step is gone
Previous versions produced the future one block at a time. By Google’s account that approach was slow, heavy on compute and vulnerable to compounding error, since each prediction rested on the output of the one before it.
TimesFM-3 instead flags every future time step as a blank and completes them all in one pass. Whether that survives long horizons in production is a different question from whether it is quicker, and speed is the claim Google is making, not stability.
One small detail carries more weight than it appears to: the output is not a lone point estimate. Nine values come back per time step, so you see the range and the uncertainty instead of a confident line that conceals how confident it really is. Anyone who has been handed a forecast with no error bars and then watched a planner treat it as scripture will recognize why that counts.
The benchmark numbers are Google’s numbers
Across Gift-Eval, FEV-Bench and Time, Google reports that TimesFM-3 takes first place among all pretrained forecasting models on both point accuracy and uncertainty calibration. Its stated competition includes Amazon’s Chronos-2, the Toto-2.0 family and Google’s own TimesFM-2.5.
Even when held to a single variable, Google says the model matches or beats that field, with the margin growing as more data is supplied. Treat the table the way you would treat any vendor’s: it belongs to the vendor.
How to get hold of it
GitHub and Hugging Face have it today. Google intends to bring it to BigQuery in the coming weeks, and that is the piece most companies will actually care about, given that TimesFM-2.5 already lives there doing single-variable forecasting via the AI.FORECAST command.
Until then, BigQuery customers remain on the one-series-at-a-time model, and the ice cream scenario is still a demo rather than a query anyone can run.
Since the family launched in 2024, Google says it has been put to work in retail, finance, manufacturing, healthcare and the sciences.
Forecasting of this kind is not the only thread being pulled. Google DeepMind put out WeatherNext Cyclones in early August, an open-source system for tropical cyclones that projects storm tracks and intensity roughly a day further ahead than the leading operational models.
If demand planning is your job and your model has been fed nothing but your own sales history, the discount schedule already sitting in your promotions calendar is the cheapest input you are ignoring. Wire that one in first.