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
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.
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.
| Model |
Params |
Weights (fp32) |
Latency |
Recommended for |
| Toto‑2.0‑4m |
4m |
16 MB |
~3.8 ms |
Edge / CPU deployment; tightest latency or memory budgets. |
| Toto‑2.0‑22m |
22m |
84 MB |
~5.0 ms |
Efficient default — matches or beats Toto 1.0 quality with ~7× fewer parameters. |
| Toto‑2.0‑313m |
313m |
1.2 GB |
~15.4 ms |
Strong general-purpose checkpoint; top-3 foundation model on GIFT-Eval. |
| Toto‑2.0‑1B |
1B |
3.9 GB |
~20.9 ms |
Best quality / cost tradeoff for production workloads. |
| Toto‑2.0‑2.5B |
2.5B |
9.1 GB |
~36.2 ms |
Highest accuracy; #1 foundation model on every benchmark. |
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
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
contiguous patch masking (CPM) for single-pass parallel decoding, a
quantile output head trained with pinball loss, a robust arcsinh input scaler, residual MLP patch projections, and is trained with NorMuon. See the
technical report for details.
Additional Resources
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},
}
