Mistral launches Timestral, an AI that forecasts sales and outages on an ordinary server

Ten milliseconds. According to Mistral AI, that is how long the smallest version of Timestral needs to forecast the next day from three weeks of measurements taken every fifteen minutes. On a plain processor, with no graphics card.

The French startup released Timestral on October 15, its first family of foundation models for time series. These are sequences of values measured over time: power consumption, daily sales, machine temperature, financial volumes. Four sizes are available, from 5 to 209 million parameters, all with the same architecture. The weights can be downloaded from Hugging Face, along with a code repository to run them on your own data.

A language model? No, a forecasting model

Timestral does not chat and does not write text. It takes the history of one or more curves and returns a probabilistic forecast: 21 quantile levels, so a range rather than a single figure. It can account for variables known in advance, such as a public holiday or a planned promotion, and accepts a history of 32,768 time steps. It works zero-shot, meaning without prior training on the customer's data.

To give an idea of scale, 200 million parameters is thousands of times fewer than large language models. That is what lets Timestral run on ordinary servers, on-premises or without an Internet connection. Mistral highlights predictive maintenance, anomaly detection on sensors, capacity and spare-parts planning, and cash management.

Scores yet to be confirmed

Mistral claims Timestral ranks first among pre-trained models on three public benchmarks: GIFT-Eval, TIME and fev-bench. According to the company, the 209-million-parameter version beats the previous state of the art with only two-thirds of its parameters. On GIFT-Eval and TIME, the 28-million version reportedly outperforms Chronos-2, Amazon's model, which is roughly four times larger.

A word of caution, though. These measurements were carried out by Mistral, which says it submitted its results to the GIFT-Eval and TIME leaderboards: we could not verify its ranking. The technical report that will detail the protocol is announced as “coming soon.” And Mistral itself acknowledges that real-world data from a factory or a store differs from benchmark data and keeps changing, which calls for fine-tuning and regular retraining.

Another point to check before any commercial use: the license. Mistral describes it as a modified MIT license, with the full text available in the repository. Pricing for any API access has not been published.

Who should try it?

If you already have an in-house forecast, for sales, service load or a site's consumption, the test is simple. Run Timestral on your own history and compare it with your current method, over a period you already know. The benefit of a model that works without a graphics card and without sending data outside the company can matter as much as a few points of accuracy.

This launch also shows that Mistral is not only playing the race for very large models, such as Mistral Large 4, whose weights are expected at the end of October. Next steps for Timestral: the technical report, and independent evaluations.