Skip to content
← Models
Open weightsDatadog/10 trustVerified

Toto

Datadog's open time-series foundation model for observability: zero-shot multivariate forecasting, 4M to 2.5B parameters.

Visit official page →

About

Toto is Datadog's foundation model for multivariate time-series forecasting, built for the spiky, high-dimensional metrics that come out of observability pipelines. The first open release (Toto-Open-Base-1.0, 151M parameters, early 2025) uses proportional factorized space-time attention with a Student-T mixture head for probabilistic forecasts, and Datadog reports top scores on both the general GIFT-Eval benchmark and its own BOOM observability benchmark. The model card on Hugging Face lists BOOM CRPS 0.375 / MASE 0.617 and GIFT-Eval CRPS 0.517 / MASE 0.75, and the checkpoint alone has over 10 million all-time downloads.

In 2026 Datadog shipped Toto 2.0, a decoder-only patched transformer scaling from 4M to 2.5B parameters with alternating time/variate attention, contiguous patch masking for single-pass long-horizon decoding, and a quantile head trained with pinball loss — plus fine-tuning and exogenous-covariate support added in February 2026. Everything is Apache 2.0. For anyone forecasting infrastructure metrics, energy load, or demand curves without wanting to train a bespoke model, Toto is the most battle-tested open option.

Specs

Context window

Variable input length (patched time steps, decoder-only)

Pricing

Free / open weights (Apache 2.0)

Free quota

Free / open weights — self-host at no license cost

API access

Hugging Face (Datadog/Toto-Open-Base-1.0, Toto-2.0 family); pip install toto-ts / toto-models; self-hosted

Weights

Open weights

Released

Not disclosed

Benchmarks

  • BOOM: CRPS 0.375, MASE 0.617 (Datadog-reported on the HF model card)
  • GIFT-Eval: CRPS 0.517, MASE 0.75 (Datadog-reported on the HF model card)
  • Toto 2.0: top spots on BOOM, GIFT-Eval, and the TIME benchmark (Datadog-reported; paper arXiv 2605.20119)

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

time-series-forecastingobservabilityanomaly-detectioncapacity-planningdemand-forecasting

Trust score — 8/10

/10 trustVerified

Scored 8/10 (verified tier, metrics-v1). Evidence: official Datadog GitHub org repo with paper (arXiv 2505.14766), BOOM dataset, and eval code; HF model card shows Apache 2.0, 151M params, 10.2M all-time downloads and self-reported BOOM/GIFT-Eval scores; Toto 2.0 (2026) scaled to 2.5B with fine-tuning support added Feb 2026. Open weights. Strengths: exceptional adoption for a TSFM, genuine open license, active maintenance, eval transparency via public leaderboards. Docked for: (1) headline benchmark numbers are Datadog-reported with no fully independent replication found; (2) optimized for observability-style series — performance on other domains is provider-claimed. License: Apache-2.0.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
9.0/10
Maintenance · recency, releases, issue load
9.0/10
Openness · license + docs
9.0/10
Manual review · sanity read + red-flag check
8.0/10
  • · Datadog/Toto-Open-Base-1.0: 151M params, Apache 2.0, 10.2M all-time downloads, 142 likes on HF
  • · BOOM CRPS 0.375 / MASE 0.617; GIFT-Eval CRPS 0.517 / MASE 0.75 (Datadog-reported on model card)
  • · Toto 2.0 (2026): decoder-only, 4M-2.5B, CPM parallel decoding, fine-tuning + exogenous covariates (Feb 2026)
  • · Official GitHub org repo with paper (arXiv 2505.14766), BOOM dataset, and eval code

Sources