Skip to main content
Back

Datadog's Toto 2.0 Finally Proves Bigger Time Series Models Get Better

Datadog AI Research proves time series foundation models scale reliably, releasing five open-weight models from 4M to 2.5B parameters that top every major forecasting benchmark

  • Scaling confirmed: Datadog AI Research proves time series foundation models scale reliably — every Toto 2.0 size strictly improves on the one below it, from 4M to 2.5B parameters.
  • State of the art on 3 benchmarks: Toto 2.0 tops BOOM, GIFT-Eval, and the contamination-resistant TIME benchmark, beating Chronos-2 and other leading TSFMs.
  • 7× more efficient than Toto 1.0: The 22M model matches Toto 1.0's quality (151M params) with 7× fewer parameters and is dramatically faster at inference via single-pass decoding.
  • No public data in pretraining: Models train only on Datadog's internal observability telemetry and synthetic data, yet lead general-purpose benchmarks — a strong test of cross-domain generalization.
  • Key innovations: Contiguous Patch Masking (parallel decoding), quantile output head, NorMuon optimizer, and u-μP hyperparameter transfer across all five model sizes from a single proxy sweep.
  • Fully open: All five checkpoints (4M, 22M, 313M, 1B, 2.5B) released under Apache 2.0 on Hugging Face; code on GitHub.

For years, one of the most stubborn open questions in time series forecasting has been whether bigger models actually get better , the way GPT-2 showed language models do. The answer, according to a new paper from Datadog AI Research, is a clear yes. The team has released Toto 2.0, a family of five open-weight forecasting models spanning 4M to 2.5B parameters, trained from a single recipe that produces reliable, monotonic quality improvements at every size.

The result is a first for the field. Time series foundation models (TSFMs) have begun to match or exceed tuned statistical baselines across heterogeneous domains, much as BERT did for language , but what TSFMs had not yet replicated from NLP and vision is reliable scaling: a single recipe applied at successively larger widths that produces predictable returns. Toto 2.0 is the first TSFM family to close that gap.

The Scaling Problem TSFMs Never Solved

In language and vision, scaling is almost boring in its reliability: train a bigger model on more data, get better results. Time series models had resisted this pattern. Competing model families often scaled unevenly, with larger versions sometimes underperforming smaller ones. The culprit was partly architectural and partly practical , hyperparameters like learning rate are not stable across model widths under standard setups, meaning teams had to re-tune from scratch for every new size, making large-scale sweeps prohibitively expensive.

The Datadog AI Research team's solution was to apply u-μP (Unit-scaled Maximal Update Parametrization) to time series for the first time. u-μP is a technique that makes the optimal learning rate independent of model width, so you can tune hyperparameters once on a small, cheap proxy model and transfer the exact same configuration to any larger model without retuning. Toto 2.0 features a family of u-μP-scaled models ranging from 4M to 2.5B parameters. The team tuned on a 10M-parameter proxy , each sweep trial completing in hours rather than days , then transferred that configuration directly to all five target sizes.

CRPS rank vs parameter count on BOOM and GIFT-Eval benchmarks, showing Toto 2.0 on the Pareto frontier at every size

What Changed Under the Hood

The backbone is a decoder-only patched transformer whose attention layers alternate between a time-axis (causal) view and a variate-axis (cross-series) view , retained from Toto 1.0. But four significant changes make Toto 2.0 a different animal: