Google has released the weights of TimesFM-3, a 330M-parameter foundation model for time series forecasting that analyzes multiple interrelated metrics together with external variables. Internally it combines patch masking with a two-dimensional attention grid, alternately attending to data over time and to connections between series. The model produces its entire forecast horizon in a single pass, so errors do not accumulate step by step and inference is faster. Trained on a trillion data points, it requires no fine-tuning for specific tasks; Google says it leads the Gift-Eval, FEV-Bench, and Time benchmarks. Weights are available on Hugging Face, with BigQuery support to follow in the coming weeks.

TimesFM-3: A Zero-Shot Foundation Model for Multivariate Forecasting