Google Research has released TimesFM-3, the third generation of its zero-shot time series foundation model. The primary update from version 2.5 is native support for multivariate inputs, allowing the model to handle multiple simultaneous targets and covariates without fine-tuning.

  • The architecture is a decoder-only transformer with 20 layers, processing 32 contiguous time steps per token using alternating causal and full attention mechanisms.
  • Forecasts are generated in a single forward pass by filling masked placeholder tokens simultaneously, outputting 9 quantiles (10th–90th percentile) per target.
  • Pretraining utilized over 1 trillion time points from sources including Wikipedia pageviews, Google Trends, and synthetic data.
  • The model claims best average rank on Gift-Eval, FEV-Bench, and Time benchmarks against Chronos-2, Toto 2.0, and TimesFM-2.5.

The weights are available under the TimesFM Non-Commercial License v1.0 on Hugging Face and GitHub, with BigQuery integration planned for the future.