Google's TimesFM-3 Breaks Time Series Forecasting

Google Research put out TimesFM-3 this week, a 330 million parameter forecasting model that breaks from every prior version of the tool. TimesFM 1.0 through 2.5 all shared one limitation: each model forecast a single time series using only that series' own past values. No context from related products, no weather data, no promotional calendars. Just one line of numbers predicting its own future.

TimesFM-3 drops that constraint. Google pretrained it on more than 1 trillion time points specifically to handle multivariate inputs, meaning it can take multiple related series, plus historical covariates, plus known future events, and forecast them together in a single pass. No fine-tuning required. The model already claims the top average rank on three separate benchmarks, GIFT-Eval, fev-bench, and the TIME leaderboard, across both point predictions and probabilistic ones.

There's a catch on deployment. The code sits under an Apache-2.0 license, but the weights themselves ship under a non-commercial license that blocks production use. Researchers can test it freely. Companies building a live forecasting product cannot ship it as-is.