Google Research has released TimesFM-3, a 330 million parameter time series foundation model that forecasts multiple related series in a single forward pass. Every TimesFM checkpoint through 2.5 was univariate: one series, its own history, nothing else. TimesFM-3 is pretrained natively for multivariate forecasting on more than 1 trillion time points, and accepts multiple targets, past covariates, and past-future covariates with no task-specific fine-tuning. It takes the top average rank among pretrained foundation models on GIFT-Eval, fev-bench, and the TIME leaderboard, on both point and probabilistic metrics.
Is it deployable?
Partial, the TimesFM repository code is Apache-2.0, but the TimesFM 3.0 weights ship under timesfm-non-commercial-license-v1.0. They are restricted to non-commercial, non-production use. You can benchmark it today. You cannot ship it behind a production forecast API.
What changed
Every TimesFM release through 2.5 was univariate. It forecast one series from its own history. Most real forecasting problems are not shaped that way. Google’s example is ice cream sales, where related product sales, foot traffic, weather, promotions, and holidays all move the target.
TimesFM-3 is pretrained natively for multivariate forecasting. It carries 330 million parameters and was pretrained on more than 1 trillion time points of real and synthetic series. Three input types work zero-shot, with no task-specific fine-tuning:
- Multiple targets forecast jointly, with point and quantile outputs for each.
- Past covariates, known only historically, such as past foot traffic.
- Past-future covariates, whose future values are known, such as a promotion calendar.
Architecture: patches, then two kinds of attention
The backbone stays a decoder-only transformer. Contiguous points are grouped into patches of 32 steps, then normalized per series so that wildly different scales do not dominate. Target and past-covariate tokens come from a single patch. Past-future covariate tokens use a lookahead trick: the current patch is concatenated with future patches, so the model sees scheduled events before they occur.
Tokens then enter a 2D grid and pass through two alternating attention mechanisms:
- Causal temporal attention runs horizontally. It is strictly causal and confined to earlier tokens inside the same series, which blocks leakage.
- Full variate attention runs vertically. At a given time step, a token reads every other series at that step, learning cross-series correlations.
One forward pass instead of many
Earlier TimesFM versions decoded one patch at a time. That adds latency, compute cost, and compounding error. TimesFM-3 uses Contiguous Patch Masking, the training-time masking strategy introduced with TiRex. Masked placeholder tokens are appended for the whole horizon. Targets and past covariates are masked there. Past-future covariates stay visible, so known future signals still reach the model. The alternating attention layers fill every masked horizon patch simultaneously. Each target receives 9 quantiles, the 10th through 90th percentile, at every horizon step.
Benchmarks
Google evaluated on GIFT-Eval, fev-bench, and the TIME leaderboard, against Chronos-2, the Toto 2.0 family, and TimesFM-2.5. Among pretrained foundation models, TimesFM-3 takes the top average rank on all three, for both point and probabilistic metrics. The package release notes record rank #1 overall on fev-bench across 100 real-world tasks, rank #1 overall on TIME across 50 domain datasets and 98 evaluation tasks, and rank #1 among foundation models on GIFT-Eval.
Interactive explainer
‘;
}
h+=’
