Google Research has released TimesFM-3, a 330 million parameter time series foundation model that forecasts multiple related series in a single forward pass. Unlike previous univariate versions, it is pretrained natively for multivariate forecasting on more than 1 trillion time points and accepts multiple targets, past covariates, and past-future covariates without task-specific fine-tuning.
- It uses alternating causal temporal and full variate attention to model cross-series dependencies zero-shot.
- Contiguous Patch Masking allows the model to produce the entire horizon in one forward pass with 9 quantiles per step.
- It ranks #1 among foundation models on GIFT-Eval, fev-bench, and TIME leaderboards for both point and probabilistic metrics.
- The weights are licensed under a non-commercial, non-production restriction; TimesFM 2.5 remains the Apache-2.0 option for shipping.
The model enables joint forecasting of multiple targets and covariates without fine-tuning, offering improved accuracy across standard benchmarks compared to prior foundation models.