Google Research has released TimesFM-3, a 330-million-parameter model that forecasts several related time series while using both historical context and known future events. A retailer, for example, can combine past ice-cream and cone sales with foot traffic, weather forecasts, holidays and scheduled discounts.
The model groups 32 consecutive values into patches and alternates attention across time and across related variables. Unlike earlier TimesFM versions that generated one future block after another, it marks the full forecast horizon as missing and fills all points in a single pass. Google says that design reduces computation and avoids errors compounding from one predicted block to the next. It produces nine values per time step to represent uncertainty rather than only one estimate.
Google trained the model on more than one trillion real and synthetic data points and says it led pretrained forecasting systems on Gift-Eval, FEV-Bench and Time. Those are company-reported benchmark results and may not transfer to every business dataset. TimesFM-3 works without task-specific training and is available through GitHub and Hugging Face. BigQuery support is planned in the coming weeks; its current forecasting command still uses the older, single-variable TimesFM-2.5.