Forecasting demand, energy use or sensor readings is the job of Timestral, a new family of models from Mistral. The company released four sizes on October 15, from 5 million to 209 million parameters.
These models work on time series, measurements arranged in time order. They can forecast a new series without first training on that particular dataset. That could reduce the setup work needed to build a separate forecasting system for each machine or product.
Timestral also accepts related inputs. Some describe the past, while others are known ahead of time, such as a holiday calendar. Its design lets forecasts use those future inputs without exposing the future values they are supposed to predict.
The largest model accepts 32,768 time steps and produces 21 quantiles. Those outputs describe a range of possible values, rather than presenting one number as certain.
Mistral reports leading results among pretrained, zero-shot models on GIFT-Eval, TIME and fev-bench. Those are company-reported evaluations. They do not establish that the model will outperform existing tools on every industrial dataset.
The company has released a repository and lists a modified MIT license. Its documentation labels the model a public preview. Mistral says a fuller technical report will follow.
For a factory or energy operator, the practical question is how well forecasts hold up when operating conditions change. The release includes local deployment and options to adapt the models to an organization's own data. The largest version has 209 million parameters.
