Google AI Releases TimesFM-3: A 330M Parameter Zero-Shot Foundation Model For Multivariate Time Series Forecasting

Details

Google Research released TimesFM-3, marking a fundamental shift in the TimesFM series. Every prior checkpoint through version 2.5 was strictly univariate — it forecast one series from its own past values. TimesFM-3 is the first version pretrained natively for multivariate forecasting, trained on over 1 trillion time points of real and synthetic data. Zero-shot, it handles three input types with no fine-tuning: multiple target series forecast jointly with point and quantile outputs; past covariates (historical-only inputs like past foot traffic); and past-future covariates (inputs whose future values are known, like a promotion calendar). On GIFT-Eval, fev-bench, and the TIME leaderboard it takes the top average rank among pretrained foundation models on both point and probabilistic metrics. The architecture keeps a decoder-only transformer backbone. Inputs are patch-tokenized (32 steps per patch) and normalized per series. Tokens pass through two alternating attention types in a 2D grid: causal temporal attention (left-to-right within a series, preventing cross-series leakage) and full variate attention (cross-series at each time step). For past-future covariates, a lookahead trick concatenates current and future patches so the model sees scheduled events before they occur. The practical catch: the repository code is Apache-2.0, but the TimesFM 3.0 weights ship under a non-commercial license. Production deployment is not permitted. Researchers can benchmark it freely; enterprises cannot ship it in production forecast APIs without a separate licensing arrangement.