Toto
Summary
Toto is Datadog’s observability-oriented time-series forecasting foundation-model line. In this wiki it covers Toto-Open-Base-1.0 and the Toto 2.0 scaling family.
Lineage
- Toto 1.0 introduces a 151M-parameter open-weights observability forecaster, BOOM benchmark, factorized time-variate attention, patch-based causal instance normalization, and Student-T mixture forecasting head.
- Toto 2.0 extends the line into an open-weights scaling family from 4M to 2.5B parameters, uses contiguous patch masking, and reports strong BOOM, GIFT-Eval, and TIME results.
Official Artifacts
- Toto source: https://github.com/DataDog/toto
- Toto-Open-Base-1.0 checkpoint: https://huggingface.co/Datadog/Toto-Open-Base-1.0
- Toto 2.0 model collection: https://huggingface.co/collections/Datadog/toto-20
- Toto 2.0 article: https://www.datadoghq.com/blog/ai/toto-2/
Role In The Wiki
Toto anchors the observability time-series branch. It is a strong passive forecasting line, but it is not yet an action-conditioned world model because deployments, rollbacks, autoscaling, remediation, and other operator actions are not first-class forecast-conditioning channels in the current sources.