Google Debuts TimesFM-3 for Multivariate Forecasting

Google Research has introduced TimesFM-3, the latest iteration of its time-series foundation model. Unlike its predecessor, TimesFM-2.5, which was released in September 2025 and limited to univariate forecasting, the new model natively supports complex multivariate scenarios. Boasting 330 million parameters, TimesFM-3 was pre-trained on a massive corpus of more than 1 trillion real-world and synthetic time points. This allows the model to capture dependencies across multiple coevolving time series without requiring task-specific fine-tuning.

The model builds on a decoder-only transformer architecture, grouping contiguous data points into patches of 32 time steps. To handle multivariate data, TimesFM-3 uses a two-dimensional grid of alternating attention mechanisms: causal temporal attention, which prevents data leakage by looking strictly backward in time, and full variate attention, which allows tokens to look across different series to learn cross-correlations. For dynamic variables like upcoming promotions or weather forecasts, the model employs a lookahead strategy that concatenates current patches with future ones.

A major shift from previous TimesFM versions is the transition to a non-autoregressive decoding strategy called Contiguous Patch Masking. Instead of generating forecasts patch-by-patch, which increases latency and compounds errors, TimesFM-3 generates the entire forecasting horizon in a single forward pass. It simultaneously predicts nine quantiles, ranging from the 10th to the 90th percentile, to provide a comprehensive probabilistic view of uncertainty.

In evaluations across three major public forecasting benchmarks—Gift-Eval, FEV-Bench, and Time—TimesFM-3 secured the top rank in both point and probabilistic forecasting metrics. It outperformed other prominent foundation models, including Chronos-2, the Toto 2.0 family, and Google's own TimesFM-2.5. The model is currently available on GitHub and Hugging Face, with integration into Google's BigQuery scheduled for the coming weeks.