Google’s forecasting family just went multivariate. TimesFM-3, the new release, packs 330M parameters and predicts several related series at once.
Earlier TimesFM releases handled one series at a time, forecasting each from its own history. The new model is pretrained natively for multivariate problems, drawing on more than 1 trillion time points of real and synthetic series, and works zero-shot across three input types without task-specific fine-tuning.
The decoder-only backbone groups points into 32-step patches, normalizing per series so different scales do not dominate. A lookahead trick lets the model see scheduled future events before they occur, while Contiguous Patch Masking fills the entire forecast horizon at once instead of decoding patch by patch. Forecasts come back as a spread: nine percentile bands, from 10th to 90th, at each step of the horizon.
In tests against Chronos-2, the Toto 2.0 family, and its own predecessor, the model took the top average rank across GIFT-Eval, fev-bench, and the TIME leaderboard, and finished first overall on fev-bench’s 100 real-world tasks.
One catch: while the code is Apache-2.0, the TimesFM-3 weights ship under a non-commercial license, so benchmarking is free but shipping the model behind a production forecasting API is not.
