Google Research has launched TimesFM-3, its latest time-series foundation model built for zero-shot multivariate forecasting. The 330-million-parameter model was pre-trained on more than one trillion real-world and synthetic time points.
Unlike earlier TimesFM versions focused on univariate forecasting, TimesFM-3 jointly analyzes multiple related time series and incorporates historical and known future variables such as promotions, weather, and holidays.
Its new architecture generates the entire forecast horizon in a single forward pass, improving efficiency and reducing error accumulation. Google reports state-of-the-art results across Gift-Eval, FEV-Bench, and Time benchmarks. TimesFM-3 is available on GitHub and Hugging Face.
