Datadog/Toto-2.0-313m

Toto-2.0-313m

Toto (Time Series Optimized Transformer for Observability) is a family of time series foundation models for multivariate forecasting developed by Datadog. Toto 2.0 is the current generation, featuring u-μP-scaled transformers ranging from 4m to 2.5B parameters, all trained from a single recipe. Forecast quality improves reliably with parameter count across the family.

The family sets a new state of the art on three forecasting benchmarks: BOOM, our observability benchmark; GIFT-Eval, the standard general-purpose benchmark; and the recent contamination-resistant TIME benchmark.

📊 Performance

<figure> <img src="assets/pareto.png" alt="Pareto frontier on BOOM and GIFT-Eval"> <figcaption>Every Toto 2.0 size sits on or near the Pareto frontier on both BOOM and GIFT-Eval. The three largest sizes rank first, second, and third among foundation models on GIFT-Eval CRPS rank. On TIME, Toto 2.0 sizes take the top three spots on every metric, ahead of every other external foundation model evaluated.</figcaption> </figure>

⚡ Quick Start

Inference code is available on GitHub.

Installation

pip install toto-models

Inference Example

import torch
from toto2 import Toto2Model

model = Toto2Model.from_pretrained("Datadog/Toto-2.0-313m")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device).eval()

# (batch, n_variates, time_steps)
target = torch.randn(1, 1, 512, device=device)
target_mask = torch.ones_like(target, dtype=torch.bool)
series_ids = torch.zeros(1, 1, dtype=torch.long, device=device)

# Returns quantiles of shape (9, batch, n_variates, horizon)
# Quantile levels: [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]
quantiles = model.forecast(
    {"target": target, "target_mask": target_mask, "series_ids": series_ids},
    horizon=96,
    decode_block_size=768,
    has_missing_values=False,
)

For more examples, see the Quick Start notebook and GluonTS integration notebook.

💾 Available Checkpoints

All five Toto 2.0 sizes share the same training recipe; pick a size based on your accuracy/latency budget. Latency is forward-pass time for a 1,024-step single-pass forecast at batch size 8 on a single A100.

✨ Key Features

• —Zero-Shot Forecasting: Forecast without fine-tuning on your specific time series.
• —Multi-Variate Support: Efficiently process multiple variables using alternating time/variate attention.
• —Probabilistic Predictions: Generate point forecasts and uncertainty estimates via a quantile output head.
• —Decoder-Only Architecture: Support for variable prediction horizons and context lengths.
• —u-μP Scaling: A single training recipe transfers cleanly across all five sizes (4m → 2.5B).

🏗️ Architecture

<figure> <img src="assets/architecture.png" alt="Overview of the Toto 2.0 architecture."> <figcaption>A decoder-only patched transformer whose attention layers alternate between time-axis (causal) and variate-axis (full) views of the input. Toto 2.0 adds <b>contiguous patch masking (CPM)</b> for single-pass parallel decoding, a <b>quantile output head</b> trained with pinball loss, a robust arcsinh input scaler, residual MLP patch projections, and is trained with NorMuon. See the <a href="https://arxiv.org/abs/2605.20119">technical report</a> for details.</figcaption> </figure>

🔗 Additional Resources

• —Technical Report
• —Blog Post
• —GitHub Repository
• —Toto 2.0 Collection — all five base checkpoints
• —BOOM Dataset — Datadog's observability time-series benchmark
• —Toto 1.0 Weights

📖 Citation

@misc{khwaja2026toto20timeseries,
      title={Toto 2.0: Time Series Forecasting Enters the Scaling Era}, 
      author={Emaad Khwaja and Chris Lettieri and Gerald Woo and Eden Belouadah and Marc Cenac and Guillaume Jarry and Enguerrand Paquin and Xunyi Zhao and Viktoriya Zhukov and Othmane Abou-Amal and Chenghao Liu and Ameet Talwalkar and David Asker},
      year={2026},
      eprint={2605.20119},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2605.20119}, 
}