Researchers have released TimesFM 3, a 330 million parameter foundation model designed for multivariate time series forecasting. The system processes related series data and planned future events in 1 single forward pass without requiring task specific fine tuning. It is available right now as open source code on GitHub and Hugging Face ahead of an upcoming BigQuery rollout.
Earlier versions of the architecture could only handle univariate numbers. This meant a system forecasting retail sales only looked at past sales figures, completely ignoring external factors like weather changes, foot traffic, or planned marketing discounts. TimesFM 3 fixes this blind spot. The new model handles multiple targets alongside both historical and future variables at the exact same time.
The internal structure relies on an alternating attention transformer. Tokens move across time steps causally to prevent future data leaks, while also communicating vertically across all data series to spot hidden connections. To cut processing delays, the model uses contiguous patch masking. Instead of generating predictions 1 patch at a time, it outputs the complete forecast horizon in 1 single computation. It also delivers 9 distinct probability quantiles ranging from the 10th to the 90th percentile to help teams measure uncertainty.
 |
| Time |
 |
| Fev-Bench |
 |
| Gift-Eval |
Training the system required a dataset of over 1 trillion synthetic and real world data points. In public benchmark testing across Gift Eval, FEV Bench, and Time, the architecture secured the top ranking among foundation models for point and probabilistic accuracy. It consistently beat competing systems including Chronos 2, the Toto 2.0 family, and earlier TimesFM releases. Even when operating in isolated univariate mode, it matched or beat existing dedicated models.
The model weights are accessible right now on GitHub and Hugging Face for public deployment. Enterprise teams will see native support added to Google Cloud BigQuery through the AI FORECAST command in the coming weeks, allowing analysts to run complex forecasts directly on data warehouse tables without machine learning expertise.
