Datadog Toto 2: Open-Source Time Series Forecasting Model Family

Signal

Datadog releases open-source Toto 2 time series forecasting model family · twitter · 2026-05-15

Datadog has released Toto 2, an open-source family of time series forecasting models ranging from 4 million to 2.5 billion parameters. The release marks the first foundational model family to empirically validate scaling laws within the time series domain, demonstrating consistent performance improvements across parameter scales.

Context

Toto 2 represents a shift toward foundational models for time series data, moving beyond traditional statistical approaches and specialized smaller neural networks. By spanning a parameter range from 4M to 2.5B, the family provides a gradient for evaluating model capacity relative to forecasting complexity. The primary technical contribution is the empirical validation of scaling laws in this domain, establishing a predictable relationship between parameter count and performance metrics. As an open-source release, Toto 2 enables integration into local and cloud-based inference stacks, supporting autonomous agents that require reliable temporal analysis without proprietary API dependencies.

Relevance

This entry expands the knowledge base's model infrastructure beyond language and vision domains, confirming that foundational model paradigms are applicable to structured temporal data. For autonomous agent workflows, Toto 2 provides a standardized, open-weight component for time series prediction, reducing reliance on closed-source forecasting APIs. The validation of scaling laws offers a benchmark for evaluating model efficiency in operational contexts, where latency and accuracy trade-offs must be quantified. It signals a maturation of open-source infrastructure where specialized domain models achieve frontier performance through scale rather than proprietary data monopolies.

Current State

Datadog has published the Toto 2 model family as open-source weights, covering parameter scales from 4M to 2.5B. The release includes validation of scaling laws, confirming performance gains with increased capacity. The models are available for integration into inference runtimes, supporting deployment on local hardware or cloud infrastructure. The ecosystem currently lacks direct counterparts in non-LLM domains, making this a distinct entry for time series foundation models.

Open Questions

• What is the composition and licensing of the training data used for Toto 2, particularly regarding proprietary Datadog telemetry?
• How does inference latency scale for the 2.5B parameter variant on consumer-grade hardware compared to smaller statistical baselines?
• Are there standardized benchmarks or evaluation suites defined for Toto 2 to facilitate cross-model comparison in the time series domain?
• How does the model handle distribution shifts and non-stationary data common in real-world monitoring environments?

Connections

No direct linkages to existing currency entries identified.