Google Research has unveiled TimesFM-3, a 330 million parameter time series foundation model designed to forecast multiple interrelated series in a single forward pass. Unlike earlier iterations—where every TimesFM checkpoint up to version 2.5 was univariate, handling one series at a time with only its own history—TimesFM-3 is natively pretrained for multivariate forecasting. It leverages over 1 trillion time points during training and accepts multiple targets, past covariates, and past-future covariates without requiring any task-specific fine-tuning.
This zero-shot capability positions TimesFM-3 as a leader among pretrained foundation models, achieving the top average rank on key benchmarks including GIFT-Eval, Fev-bench, and the TIME leaderboard, across both point and probabilistic metrics.
Architecture and Capabilities
The model’s architecture enables it to process diverse input types seamlessly, making it highly versatile for real-world forecasting tasks. By eliminating the need for fine-tuning, TimesFM-3 simplifies deployment for practitioners who require robust, out-of-the-box predictions.
Performance and Benchmarks
TimesFM-3 outperforms other pretrained models on multiple evaluation suites, demonstrating consistent superiority in accuracy and reliability. Its top average rank underscores its effectiveness across varied datasets and forecasting horizons.
Availability and Future Implications
As of late 2026, TimesFM-3 is openly available, allowing researchers and developers to integrate it into their workflows. For more details, refer to the original blog post on Google Research.
via MarkTechPost