Mistral releases Timestral, open time series models that run on a CPU
Mistral has released Timestral, its first family of time series foundation models, as open models. Four models from 5 million to 209 million parameters make probabilistic, multivariate forecasts without a GPU.
On October 15, 2026, Mistral released Timestral, its first family of time series foundation models. The models look at past values of a sensor, sales figures or energy use and forecast what comes next, and they are offered as open models you can run on your own servers. The code is on GitHub and the models are on Hugging Face.
Four sizes, no GPU needed
The family has four models with 5 million, 28 million, 123 million and 209 million parameters. According to Mistral, Timestral does not need a GPU to forecast: on an 8-thread server CPU, a one-day-ahead forecast takes 10 milliseconds with the smallest model and 140 milliseconds with the largest. That makes it easier to run on a factory PC or an air-gapped network.
A range of outcomes instead of one number
Instead of a single number, Timestral gives probabilistic forecasts with 21 quantiles, showing optimistic and pessimistic scenarios next to the most likely value. It can forecast several series together and use extra information that shapes the forecast: values known only for the past, future values known in advance such as holidays or planned campaigns, and categorical data. The context length is 32,768 time steps. Mistral lists sensor telemetry, equipment drift, maintenance and spare-parts planning, financial volumes and supply chains among the uses, and shows a workflow where Timestral flags an anomaly and agents start a root-cause analysis.
License and deployment
The models ship under a "Modified MIT" license. Organizations can deploy them on-premises, in a private cloud or air-gapped through Mistral Forge, and fine-tune them on their own data.
Mistral says Timestral ranks first among pretrained zero-shot models on all five metrics across GIFT-Eval, TIME and fev-bench, and that the 28-million-parameter model beats Chronos-2, which is about four times larger. The full technical report has not been published yet, and the results so far rest on charts shared by Mistral. It is worth testing the models on your own history next to your current forecasting method and watching for comparisons from independent researchers.
