Mistral launches Timestral to forecast time series

Mistral has introduced Timestral, a family of AI models designed to forecast how data changes over time, from industrial sensors and energy consumption to sales, traffic, and financial movements.

Its largest model has 209 million parameters, but the company says it outperforms larger earlier models on the sector's three main public benchmarks: GIFT-Eval, TIME, and fev-bench. The results have been submitted to those leaderboards and come from a single model, without combining multiple predictions.

More useful forecasts for businesses

A time series is a sequence of data ordered by time. For example, a machine's temperature every 15 minutes, a store's daily order count, or a building's electricity consumption.

Timestral does more than calculate a future value. It also provides probabilistic forecasts, with 21 possible quantiles showing the range of expected outcomes. This means a company can estimate not only that demand will probably reach a certain level tomorrow, but also how likely it is to be much higher or lower.

The system can also work with several related time series at once. In a factory, it could analyze a pump's vibration, pressure, and temperature together instead of treating each signal as isolated.

It also accepts additional information that may affect the forecast:

• Past data, such as sales history or sensor readings.
• Information known in advance, such as a production schedule, a planned shutdown, or a weather forecast.
• Categorical data, such as whether a machine is running, stopped, or under maintenance.

This makes it possible, for example, to anticipate an increase in load because it was scheduled, rather than interpreting it as an unexpected anomaly.

Four sizes and CPU operation

Timestral will be available in four versions: 5 million, 28 million, 123 million, and 209 million parameters. The smaller models trade some accuracy for lower resource use and faster responses.

Mistral says all four offer a good balance between size and accuracy compared with comparable public models. According to its tests, the 28-million-parameter model outperforms Chronos-2 on GIFT-Eval and TIME despite having approximately one-quarter as many parameters.

You do not need a GPU to use Timestral. On a conventional server with eight CPU threads, generating a one-day forecast from three weeks of data recorded every 15 minutes takes 10 milliseconds with the 5-million-parameter model and 140 milliseconds with the 209-million-parameter model.

The family supports contexts of up to 32,768 time steps and can be deployed on your own servers, private clouds, or environments isolated from the internet.

From an alert to an explanation

Mistral's idea is for Timestral to become part of broader systems. In an industrial plant, the model could monitor thousands of sensors and detect that a pump is behaving abnormally.

Then, another AI agent could compare those signals with historical failures, review maintenance reports, consult manuals, and analyze operators' notes. The result would be a suggested diagnosis with possible causes and next steps, not just an isolated alert.

That workflow still requires several systems to be integrated and its decisions to be validated. Timestral is the numerical model that generates forecasts, not an automatic replacement for the staff who must decide whether to stop a machine or schedule a repair.

A model trained first on synthetic data

Mistral explains that most of the training used synthetic data, meaning artificially created time series designed to represent trends, seasonality, noise, and relationships between variables. It then combined that data with real-world information in a second training phase.

The company also offers adaptation through Mistral Forge. Organizations can fine-tune the model with their own data, add variables specific to their industry, and retrain it when demand patterns, equipment behavior, or market conditions change.

The repository for trying Timestral is already available under a modified MIT license. Mistral has said it will publish a technical report with the complete results later.

For you, the important point is that AI models are starting to cover a task that is less visible than chatbots but widespread in business: anticipating what will happen in day-to-day operations. The question to watch now is whether the announced benchmark advantage holds when Timestral faces real-world data that is changing and far less orderly.