Mistral AI has released Timestral models for time series forecasting
Mistral AI has released Timestral, its first family of models for time series forecasting. The four variants have 5 to 209 million parameters and handle multiple related series as well as covariates. According to the company, they do not require a GPU to run.
Mistral AI has released the Timestral family in four sizes: 5, 28, 123 and 209 million parameters. The models produce probabilistic forecasts for multiple related time series and support both historical and known future covariates—additional variables such as a production schedule. According to the company, they can also forecast from previously unseen data in a zero-shot setting, meaning without fine-tuning for a specific task.
According to company measurements, forecasting one day from three weeks of history at 15-minute intervals on a server CPU with eight threads takes 10 ms for the smallest variant and 140 ms for the largest. The company says the models do not need a GPU to compute forecasts.
According to the company’s evaluation, the variant with 209 million parameters ranks first on GIFT-Eval, TIME and fev-bench among pretrained models in a zero-shot setting. The results come from a single model checkpoint without combining multiple models. The company also says domain-specific fine-tuning is available through the Mistral Forge product.
Why it matters
A forecast accompanied by an uncertainty range can support risk-based decisions about capacity or maintenance. Accounting for planned shutdowns or production changes also makes it possible to incorporate events known in advance when estimating future developments.
Two audiences, two different impacts
What this means
For individuals
Based on the described capabilities, analysts can try forecasting their own time series without training the model first and test whether covariates help.
For a business
For companies with existing monitoring systems, the claimed ability to run on a CPU offers a way to add forecasts without a dedicated GPU. However, infrastructure requirements depend on model size and the specific workload.
ProcessesCheck the original
Event sources
clearly official source · 1 publisher, 0 independent. We count feeds from the same owner only once.