Google's TimesFM 3.0 Forecasts Multiple Time Series at Once Without Retraining

Google Research released TimesFM 3.0, a 330-million-parameter forecasting model that can predict several related time series simultaneously, incorporating signals you already know about the future, all without task-specific fine-tuning. Previous versions handled only one series at a time. This third generation is the first pretrained natively for multivariate forecasting, letting related signals inform each other in a single forward pass.

There is a licensing catch worth knowing upfront. TimesFM 2.5 shipped under Apache 2.0, but the 3.0 weights carry a new timesfm-non-commercial-license-v1.0 that explicitly prohibits production deployment and third-party mirroring. Benchmarking, prototyping, and academic publication are permitted. Shipping it behind a customer-facing forecasting API is not.

What changed under the hood

The architecture uses alternating causal and cross-series attention layers across 20 transformer layers at model dimension 1280 with 16 heads, a 32-step context patch length, and a 64-patch forecast horizon, trained on over a trillion data points. The alternating attention design is what enables zero-shot multivariate forecasting: causal layers handle temporal dependencies within a series, while cross-series layers let signals inform each other.

Inference is also restructured. Rather than decoding patch by patch, the model generates the entire forecast horizon in one forward pass using Contiguous Patch Masking, a training-time strategy introduced with TiRex. Masked placeholder tokens are appended for the full horizon. Targets and past covariates are masked there, while past-future covariates remain visible, so known future signals still reach the model. Every masked horizon patch is filled simultaneously.