Toto-2.0: The Upgraded Time-Series Foundation Model (+Retail Forecasting Tutorial)

It feels like every few weeks, there’s a breakthrough in time-series foundation models.

When Datadog Research released Toto 1.0, it was a massive step for observability forecasting. Now, they are back with a generational leap: Toto 2.0.

For a while now, the forecasting community (myself included) has been debating whether the scaling laws that power LLMs actually apply to time series. Does increasing the parameter count of a model automatically make it a better forecaster? Toto 2.0 was designed to answer exactly that. Ranging from 4M up to 2.5B parameters, the results show that scaling works for time series, too.

Beyond the scaling, Toto 2.0 is significantly faster and more parameter-efficient than its predecessor. It currently takes the top spot on major benchmarks like BOOM, GIFT-Eval, and the new data-leakage-free TIME benchmark.

In this article, we’ll walk through an extensive benchmark, demonstrating how to use Toto 2.0 on a retail forecasting case. In a follow-up article, we'll dive into the architecture and see exactly how Toto-2.0 works under the hood.

Let’s get started!

✅ Find my Toto 2.0 use case project in the AI Projects folder (Project 29) — plus more cool projects inside!

Toto 2.0 is Datadog’s decoder-only foundation model built for multivariate zero-shot time-series forecasting.

Key features:

• 𝗗𝗲𝗰𝗼𝗱𝗲𝗿-𝗼𝗻𝗹𝘆 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲: It utilizes a patched transformer design that alternates between time-axis and variate-axis attention (like Toto-1). It natively supports variable context lengths and accepts future known variables (exogenous covariates).

• 𝗜𝗻𝗻𝗼𝘃𝗮𝘁𝗶𝘃𝗲 𝘂-μ𝗣 𝗦𝗰𝗮𝗹𝗶𝗻𝗴: Training massive models usually requires expensive, trial-and-error hyperparameter tuning at every scale. By adopting u-μP (Maximal Update Parametrization) from the LLM research, Toto 2.0 bypasses this bottleneck entirely. The learning dynamics are optimized once on a lightweight 4M-parameter model, and those exact hyperparameters transfer perfectly to the 2.5B-parameter version, delivering clean and predictable scaling.

• 𝗡𝗲𝗮𝗿-𝗶𝗻𝘀𝘁𝗮𝗻𝘁 𝗹𝗮𝘁𝗲𝗻𝗰𝘆 𝘄𝗶𝘁𝗵 𝗖𝗣𝗠: Toto 2.0 integrates Contiguous Patch Masking (CPM)—a training strategy introduced by the TiRex model. Instead of relying on slow, step-by-step autoregressive generation, this technique enables single-pass parallel decoding, allowing the model to generate long horizons (up to 1,024 steps) with near-instant latency.

• 𝗣𝗿𝗼𝗯𝗮𝗯𝗶𝗹𝗶𝘀𝘁𝗶𝗰 𝗳𝗼𝗿𝗲𝗰𝗮𝘀𝘁𝗶𝗻𝗴: It generates both precise point forecasts and reliable uncertainty estimates using a quantile output head trained with pinball loss.

• 𝗦𝘂𝗽𝗲𝗿𝗶𝗼𝗿 𝗯𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸 𝗿𝗲𝘀𝘂𝗹𝘁𝘀: It achieves state-of-the-art performance across major evaluations, outperforming previous models on both the GIFT-Eval and the new TIME benchmarks. Furthermore, Toto 2.0 is an open-weight model released under the Apache-2 license.

Next, we’ll put Toto 2.0 to the test in a practical, Retail-Forecasting benchmark against strong statistical models.