Google's TimesFM-3 forecasts multiple series at once, but its license blocks production use

Google Research just dropped TimesFM-3, and the headline feature is a fundamental shift: every prior TimesFM checkpoint through 2.5 was univariate, forecasting one series from its own history. TimesFM-3 is natively multivariate. It carries 330 million parameters and was pretrained on more than 1 trillion time points of real and synthetic data. The model accepts multiple targets, past covariates, and past-future covariates zero-shot, with no task-specific fine-tuning. Google’s go-to example is ice cream sales, where product sales, foot traffic, weather, promotions, and holidays all move the target together.

The architecture stays a decoder-only transformer, but with a twist. Contiguous points get grouped into patches of 32 steps, then normalized per series so different scales don’t dominate. Tokens then enter a 2D grid with two alternating attention mechanisms: causal temporal attention runs horizontally within a series, while full variate attention runs vertically across all series at a given time step. Past-future covariates use a lookahead trick where the current patch is concatenated with future patches, so the model sees scheduled events like promotions before they occur.

Latency was a real problem in earlier versions. TimesFM decoded one patch at a time, adding compute cost and compounding error. TimesFM-3 fixes this with Contiguous Patch Masking, the training-time masking strategy introduced with TiRex. Masked placeholder tokens are appended for the whole horizon, and the alternating attention layers fill every masked patch simultaneously. Each target gets 9 quantiles, from the 10th through 90th percentile, at every horizon step.

On benchmarks, TimesFM-3 takes the top average rank among pretrained foundation models on GIFT-Eval, fev-bench, and the TIME leaderboard, for both point and probabilistic metrics. The release notes record rank #1 overall on fev-bench across 100 real-world tasks and rank #1 overall on TIME across 50 domain datasets and 98 evaluation tasks. That’s a meaningful jump over Chronos-2, the Toto 2.0 family, and TimesFM-2.5. But here’s the catch that will frustrate plenty of practitioners: the repository code is Apache-2.0, but the TimesFM 3.0 weights ship under a non-commercial, non-production license. You can benchmark it today. You cannot ship it behind a production forecast API. TimesFM 2.5 remains the Apache-2.0 option if you actually need to deploy.