Google Releases TimesFM-3, a 330M-Parameter Multivariate Forecasting Model

Google Research released TimesFM-3, a 330 million parameter time series foundation model that forecasts multiple related series in a single forward pass. [1] The model is pretrained natively for multivariate forecasting on more than 1 trillion time points. [1] It accepts multiple targets, past covariates, and past-future covariates with no task-specific fine-tuning. [1] Previous TimesFM checkpoints through 2.5 were univariate, meaning they forecasted one series solely from its own history. [1]

TimesFM-3 achieves the top average rank among pretrained foundation models on GIFT-Eval, fev-bench, and the TIME leaderboard across both point and probabilistic metrics. [1] While the repository code uses an Apache-2.0 license, the model weights ship under a restrictive non-commercial license. [1] This means users can benchmark the model today, but they cannot ship it behind a production forecast API. [1]

The shift from univariate to multivariate forecasting allows TimesFM-3 to handle complex, real-world scenarios where variables interact, such as retail sales impacted by weather and promotions. By topping benchmarks like the TIME leaderboard zero-shot, TimesFM-3 demonstrates the viability of large-scale foundation models for complex forecasting previously requiring bespoke fine-tuning. However, the restrictive non-commercial license prevents immediate enterprise deployment, leaving organizations to rely on other open models for production APIs. What the sources don't address: whether Google intends to release a commercially permissive version of the TimesFM-3 weights in the future.