Toto 2.0 Sets New Standards in Time Series Forecasting with Scalable Foundation Models

The researchers demonstrate that time series foundation models can effectively scale, achieving reliable forecast-quality improvements ranging from 4 million to 2.5 billion parameters. They introduce Toto 2.0, a family of five open-weights forecasting models developed using a consistent training methodology. Toto 2.0 establishes a new state of the art across three forecasting benchmarks: BOOM, GIFT-Eval, and the contamination-resistant TIME benchmark. This report details the experimental results, design choices, and architecture of Toto 2.0, as well as the training data and the u-muP hyperparameter transfer pipeline. All five base checkpoints are made available under the Apache 2.0 license.