Datadog Toto 2 0 2 5b

Datadog Toto 2.0 2.5B — 2.45B-parameter decoder-only transformer for zero-shot multivariate time series forecasting. Trained with u-μP scaling, alternating time/variate attention, contiguous patch masking (CPM) for single-pass parallel decoding, and a quantile output head (0.1–0.9). Highest-accuracy checkpoint in the Toto 2.0 family (4m → 2.5B). #1 foundation model on BOOM, GIFT-Eval, and TIME benchmarks. Supports variable context/horizon, missing values via target_mask, and GluonTS integration. Requires `toto-models` (Python 3.12+, PyTorch 2.5+). Use when the user needs high-accuracy zero-shot multivariate time series forecasting, observability metric forecasting, or spe…