TimesFM-3 introduces a zero-shot foundation model for multivariate time series forecasting, supporting multiple targets, past and dynamic covariates, and joint prediction of coevolving series. With 330 million parameters pre-trained on over 1 trillion time points, it outperforms prior univariate models across benchmarks in retail, finance, healthcare, and other domains. The decoder‑only transformer uses patch‑based tokenization, per‑series normalization, and an alternating attention architecture that combines causal temporal attention with full variate attention. This enables a single‑pass, non‑autoregressive decode, improving efficiency and accuracy while handling complex real‑world scenarios such as forecasting ice cream sales with auxiliary data like promotions and weather. Times FM-3 marks a shift from strictly univariate to native multivariate forecasting, expanding applicability and adoption in production environments.

Key innovations include support for multiple targets (point and quantile forecasts), incorporation of past and future covariates via a lookahead token strategy, and a 2D attention grid that captures both temporal dynamics and cross‑series correlations. By processing patches of 32 time steps and using a contiguous patch masking technique, the model achieves zero‑shot generalization without task‑specific fine‑tuning, delivering state‑of‑the‑art performance on major forecasting benchmarks.

The release builds on the success of Times FM and Times FM‑2.5, extending Google AI’s time‑series foundation model lineage to address the inherently multivariate nature of real‑world forecasting problems. This advancement is expected to accelerate adoption across diverse industries, from retail and finance to manufacturing and natural sciences.

330 million, 1 trillion, Times FM, Times FM‑2.5, Google AI