Google has released TimesFM-3, the latest generation of its time-series foundation model and the first in the series to support native multivariate forecasting. The 330-million-parameter model is pretrained on a corpus of real-world and synthetic data exceeding one trillion time points.

TimesFM-3 can predict multiple interrelated time series simultaneously in a single forward pass, without task-specific fine-tuning, while preserving the zero-shot generalization that characterized earlier versions. Previous models in the family, including TimesFM-2.5, which was released in September 2025, were strictly univariate and forecast individual series using only their own historical values.

Google said TimesFM-3 is designed for the way most real-world forecasting problems work, where multiple coevolving signals must be considered together. The model supports joint prediction of several target variables, the use of historically observed covariates such as past foot traffic, and future-known dynamic inputs including promotional schedules and weather forecasts.

Under the hood, TimesFM-3 uses a decoder-only transformer backbone and processes data in contiguous patches of 32 timesteps with per-series normalization. Its main innovation is an alternating attention mechanism that operates across a two-dimensional token grid. Causal temporal attention ensures that each token can access only past data within its own series, helping prevent information leakage, while full variate attention allows the model to learn correlations across series at each timestep.

For future-known covariates, the system concatenates current and upcoming patches into lookahead tokens. Google also said the move to non-autoregressive inference through Contiguous Patch Masking allows the model to generate the full forecast horizon at once, removing the latency and error accumulation associated with patch-by-patch generation in earlier versions.

Introducing TimesFM-3, a state-of-the-art time series foundation model that enables accurate multivariate time series forecasting in a single forward pass, significantly outperforming other forecasting models across major benchmarks. More on the blog → pic.twitter.com/NfbONpFYDz — Google Research (@GoogleResearch) August 31, 2026

Google Research repeated the announcement in a post on X, writing: "Introducing TimesFM-3, a state-of-the-art time series foundation model that enables accurate multivariate time series forecasting in a single forward pass, significantly outperforming other forecasting models across major benchmarks. More on the blog → pic.twitter.com/NfbONpFYDz" — Google Research (@GoogleResearch) August 31, 2026

Benchmark Performance and Cloud Integration

Google evaluated TimesFM-3 on three public benchmarks: Gift-Eval, FEV-Bench, and Time. The company said the model achieved state-of-the-art results in both point and probabilistic forecasting metrics among pretrained foundation models. The comparisons included multivariate-capable models such as Chronos-2 and the Toto 2.0 family, as well as the earlier TimesFM-2.5.

Google said that even when TimesFM-3 was limited to univariate mode without cross-series or covariate information, it matched or exceeded competing models. When full multivariate operation was enabled, the model delivered additional gains by using inter-series dependencies.

The model outputs nine quantiles from the 10th to the 90th percentile, providing a detailed uncertainty profile rather than a single forecast value.

For enterprise use, TimesFM-3 is available on GitHub and Hugging Face. Google said BigQuery integration will arrive in the coming weeks, extending the AI.FORECAST command introduced with TimesFM-2.5 to multivariate workloads. That matters for teams that already keep operational data in BigQuery, because it points to forecasting inside existing analytics workflows rather than in a separate tool. Google said this will allow organizations in retail, finance, manufacturing, and scientific research to incorporate external signals such as planned promotions directly into forecasting pipelines, so they can account for demand fluctuations rather than only extrapolating historical patterns.