Try Timestral against a simple forecast first

Short answer

Mistral released Timestral on October 15, 2026, in four sizes from 5M to 209M parameters. Use the official model IDs below. The downloadable example runs simple baselines by default. Timestral inference is optional and was not executed for this guide.

Which Timestral model should you open?

Timestral is Mistral’s time-series forecasting family. Zero-shot forecasting means using a pretrained model without first fitting it to your own series. It still needs historical observations. Official release.

These are parameter-size variants of the same forecasting family. The model cards describe univariate and multivariate inputs, missing-value handling, and probabilistic outputs. A quantile is a predicted value at a selected probability level. Main model card.

For a small first trial, our example selects mistralai/Timestral-5M. This is a resource-conscious starting point, not a claim that it is the best model for your data.

Check the runtime and the separate licenses

The official Python package requires Python 3.10 or later and PyTorch 2.4 or later. Mistral documents CPU and GPU inference. Our optional example selects CPU and one PyTorch thread. Available memory and total runtime still depend on your environment. Pinned setup documentation.

The repository code is Apache-2.0 licensed. Model weights have a separate Modified MIT license. The checked weight license has a revenue condition. The restriction applies if your company’s or employer’s global consolidated monthly revenue exceeded $20 million in the preceding month. The threshold covers the equivalent amount in another currency. The condition also applies to derivatives and combined works. Read the complete terms before using a checkpoint. 5M weight license.

The local example does not call a paid hosted API. Model files, package installation, disk use, and computation still consume resources. Hosted providers can have separate prices and terms.

Run the baseline check without downloading Timestral

Download compare_forecast.py. Save it in an empty working folder. The default mode uses only Python’s standard library.

python3 compare_forecast.py

The command saves timestral-result.json in your current folder. Read that file with the next command.

python3 -m json.tool timestral-result.json

For a separate run, choose a new output filename.

python3 compare_forecast.py --output second-baseline-result.json

The script creates 63 invented daily observations with an exactly repeating weekly pattern. It gives the first 56 values to the prediction methods. It keeps the last seven values separate for scoring.

We ran this default mode. Its mean absolute error, or MAE, was 6.0 for repeating the last value and 0.0 for repeating the previous week. MAE is the average absolute difference between predictions and the held-back values. Lower is better on the same series and holdout.

The zero error is expected from this deliberately simple pattern. It is a check of the example’s arithmetic and split. It says nothing about Timestral’s accuracy on real demand or sensor data. The saved result reports model_executed: false.

Run Timestral only after checking the setup

Use a separate environment with Python 3.10 or later. Git must be available for this source-based installation. Check that your platform supports the required Python and PyTorch versions. Install the official package at the source revision checked for this article.

git --version
python3 --version
python3 -m venv .venv
source .venv/bin/activate
python -m pip install "git+https://github.com/mistralai/timestral.git@a534bdbd244d64f3d9d743f46c7bf6fc330ed4ff"

These activation commands target macOS and Linux shells. On Windows, use the activation command for your environment. Package installation and model inference were not executed in our verification.

python compare_forecast.py --run-model --output timestral-model-result.json

This flag downloads the public 5M checkpoint if it is absent from the cache. The example pins its model revision separately from the package revision. It passes only the 56 context values to the model, then scores the median prediction against the seven held-back values.

The script saves a JSON result with the model ID, revision, and scores. It refuses to overwrite an existing output file. Read a saved result before starting another run. An installation error, missing result, or failed inference is not a completed forecast.

Use your own history without giving away the answer

Keep dates in order. Use one consistent sampling interval. Choose a forecast horizon that matches the decision you need to make. These are suggested evaluation steps, not additional Timestral API requirements.

Separate the future target values before constructing model inputs. A known-future covariate is an extra input available at forecast time, such as a planned schedule. Later observations must not enter that input. Otherwise, the test gives the model information it would not have had.

On your own data, repeat the comparison at several historical cutoffs. Keep the units, horizon, scoring method, and input history consistent across methods. Record failures and inference time alongside prediction error. A single synthetic sequence cannot establish production value.

An assistant can prepare the data split and a report for review. Direct Timestral integration with OpenAI Dots was not verified. Use the first-task guidance to define the input and expected result.

FAQ