Toto 2.0: Time series forecasting enters the scaling era (www.datadoghq.com)

The release of Toto 2.0, a new family of open-weight time series forecasting models, marks a significant advancement in the AI/ML community. Spanning from 4 million to 2.5 billion parameters, Toto 2.0 demonstrates that time series foundation models (TSFMs) can increase in effectiveness as they scale—yielding best-in-class results across multiple benchmarks including BOOM, GIFT-Eval, and TIME. Notably, Toto 2.0 is seven times more parameter-efficient than its predecessor, Toto 1.0, and offers enhanced inference speed with innovations like contiguous patch masking (CPM) that enable parallel forecasting. This model is trained on observability and synthetic data, yet manages to generalize effectively to new datasets. Toto 2.0 not only achieves substantial improvements in forecast quality but also sets a precedent for the future of TSFMs. Its scalable architecture aligns with findings in other domains, showing that larger models reliably outperform smaller counterparts—a milestone previously unverified in time series forecasting. The implications extend beyond mere performance metrics; there's a pressing need for improved data curation and evaluation tailored to the unique complexities of time series data. As Toto 2.0 reshapes expectations and standards for forecasting models, it highlights the evolving landscape of AI/ML and opens up new avenues for research and application in time series analysis.