One-seventh of the parameters outperform the previous generation, pretraining can achieve cross-domain generalization using only monitoring metrics and synthetic data—data efficiency surprises me more than model size.

Datadog announces the open-source time series forecasting model Toto family, with five versions: 4m, 22m, 313m, 1B, 2.5B, all under Apache 2.0. Toto 2 is the first to verify the scaling law in the time series domain, the larger the scale, the stronger the prediction, with 2.5B still unsaturated; it wins on BOOM, GIFT-Eval, and TIME benchmarks. Introducing continuous graph block masking, converting autoregression to unidirectional feedforward, significantly speeds up, with 313m latency approaching Chronos-2's 120m. Pretraining only uses system monitoring metrics and synthetic data, yet still demonstrates cross-domain generalization; the 22m version beats Toto 1.0 with only one-seventh of the parameters.