---
title: "Chronos-2 vs. Timestral: Load Forecast Backtest | alitiq"
description: "We tested Chronos-2 and the new Timestral zero-shot on the day-ahead electricity load of a Swiss town. Both reach an MAE below 5 %. Results and limits."
url: https://alitiq.com/en/blog/foundation-models-load-forecasting-chronos-timestral
language: en
alternates:
  en: https://alitiq.com/en/blog/foundation-models-load-forecasting-chronos-timestral
  de: https://alitiq.com/de/blog/foundation-models-lastprognose-chronos-timestral
---

[Back to Blog](https://alitiq.com/en/blog)

Load Forecast

Foundation Models

AI

# Chronos-2 vs. Timestral: Foundation Models for Load Forecasting

October 17, 2026

Daniel Gröger

![Chronos-2 vs. Timestral: Foundation Models for Load Forecasting](https://alitiq.com/images/products/lastprognose-stromnetz-stadt-illustration.webp)

Day-ahead load forecasts for a municipal grid

On 15 October 2026, Mistral AI released Timestral, a family of time-series foundation models. We wanted to know how it does on a task our customers care about: forecasting the electricity load of a town one day ahead. So we put it side by side with Chronos-2 from Amazon, one of the established open models in this class.

## What a time-series foundation model does

A language model reads a prompt and writes what comes next. A time-series foundation model does the same with numbers: it reads recent history and continues it.

- **Pre-trained once** on large collections of real and synthetic time series.
- **Given context** at forecast time: recent load plus weather and calendar covariates.
- **Forecasts zero-shot:** out comes a probabilistic forecast, without fitting a model for the site.

That is the appeal: no training pipeline per site, and a forecast from the first day.

## The two models

|  | Chronos-2 | Timestral-209M |
| --- | --- | --- |
| Developer | Amazon | Mistral AI |
| Parameters | 120 M | 209 M |
| Context window | 8,192 time steps | 32,768 time steps |
| Covariates | past and future | past and future |

## The test setup

- **Target:** electricity load of a Swiss town, forecast day-ahead in quarter-hour resolution.
- **Covariates:** several weather and calendar features, passed to both models as extra inputs.
- **Period:** one month, 2 March to 1 April 2025.
- **Zero-shot:** neither model was trained or fine-tuned on the site.

## Results

Both models reach a day-ahead MAE below 5 % of the mean load. Timestral is ahead on all three metrics:

| Model | MAE | RMSE | BIAS |
| --- | --- | --- | --- |
| Chronos-2 | 4.79 % · 0.87 MW | 6.27 % · 1.14 MW | −0.37 MW |
| Timestral-209M | 4.12 % · 0.75 MW | 5.59 % · 1.02 MW | −0.17 MW |

![One week from the backtest: measured load and the day-ahead forecasts of Timestral-209M and Chronos-2, 10 to 16 March 2025](https://alitiq.com/images/blog/load-forecast-foundation-models-chronos-timestral.webp)

One week from the backtest: measured load and the day-ahead forecasts of Timestral-209M and Chronos-2, 10 to 16 March 2025

The week above shows what the numbers mean in practice. Both models follow the daily cycle and the lower weekend load. The errors sit mainly in the peaks and in short dips that neither model sees coming.

MAE and RMSE are given in MW and as a percentage of the mean load. BIAS is measurement minus forecast: the negative values mean both models forecast slightly too much load on average, Timestral by less than half as much as Chronos-2. Why we report errors relative to the mean load rather than as MAPE is explained in our article on [the pitfalls of MAPE](https://alitiq.com/en/blog/mape-pitfalls-forecast-accuracy).

For models that have never seen this town's data, that is a strong baseline.

## Where the limits are

**The context window is finite.** A foundation model only sees what fits in its window. At 15-minute resolution, Chronos-2's 8,192 steps cover about 85 days and Timestral's 32,768 steps about 341 days. Last year's holidays or the same season a year ago fall outside that window. Our production load models learn from about three years of history.

**Compute is not free.** Running this one-month backtest took about two hours on a laptop. That is fine for a test, but it adds up for daily forecasts and regular backtests across a portfolio of many sites.

**One site, one month.** The test covers a single town in March. It is a first look, not a general verdict: other load profiles and seasons can rank the models differently.

## What we take from it

- **Strong baseline:** zero-shot day-ahead MAE below 5 %, with no model trained for the site.
- **Context is finite:** the window caps how much history a forecast sees, and backtests need real compute.
- **Europe competes:** with Timestral, an open European model now meets Chronos-2 head to head.

A strong European competitor in time-series AI is good news for the energy sector. Timestral comes in four sizes from 5 M to 209 M parameters, runs on CPU without a GPU and is available with open weights on Hugging Face. Merci, Mistral AI.

If you want to see how foundation models and our production [load forecasts](https://alitiq.com/en/load-forecast) perform on your own data, talk to us.

### Ready to Transform Your Business?

Discover how alitiq's forecasting intelligence can optimize your operations.

[Contact Us](https://outlook.office.com/bookwithme/user/a4a08a92ad764515bccc3a29740b96ff%40alitiq.com?anonymous&ismsaljsauthenabled=true)