TabTalk #4: Toto 2.0 by Datadog

Do time series foundation models (TSFMs) scale reliably? In this session, Gerald Woo presents Toto 2.0, a newly released family of open-weights forecasting models across five sizes (4M, 22M, 313M, 1B, and 2.5B parameters) that demonstrates clear, monotonic improvement with scale. Gerald will dive into the core innovations driving Toto 2.0's state-of-the-art performance across major benchmarks (GIFT-Eval, TIME, and BOOM). He will detail model architecture updates from Toto 1.0 to Toto 2.0, the training data mix comprising of only observability and synthetic data, and the scaling recipe used to achieve reliable scaling results.

‍

Gerald Woo is a Senior Research Scientist at Datadog AI Research. He has pioneered work on time series foundation models and continues to push the boundaries at Datadog. Before this, he obtained his PhD at Singapore Management University and Salesforce AI Research under the Industrial Postgraduate Programme.