Mistral releases Timestral, a new family of time-series forecasting models
The French AI lab says its smallest model tops every major benchmark while running on a standard CPU.
Mistral has unveiled Timestral, its first family of time-series foundation models, designed to predict future values from sequences of past data such as sensor readings, energy use, or financial flows.
A time series is simply a list of numbers collected in order over time. Forecasting means using that history — and sometimes outside information — to predict what comes next.
The largest Timestral model has just 209 million parameters 1. A parameter is a number the model adjusts during training so it can learn patterns. Despite its small size, Mistral says it ranks first on all three major public forecasting benchmarks: GIFT-Eval, TIME, and fev-bench, beating previous top models while using only about two-thirds as many parameters.
The family includes four sizes — 5 million, 28 million, 123 million, and 209 million — so users can choose between speed and accuracy. The 5M model needs only 10 milliseconds to forecast one day ahead from three weeks of 15-minute sensor data on a regular server CPU, while the 209M version takes 140 milliseconds 1. None of them require a GPU, the specialized graphics chip many AI models rely on.
Mistral trained Timestral on a large synthetic dataset — data created by computers rather than collected from the real world — so it can make zero-shot forecasts, meaning predictions on data it has never seen before, even in fields where real examples are scarce.
The model can also look at related series at the same time. If a pump's vibration, temperature, and pressure all change together, Timestral can use those links instead of guessing each one on its own. It handles categorical inputs — labels like "idle" or "under maintenance" that do not have a natural order — and known-future inputs such as scheduled production changes, so a planned load increase does not look like a surprise.
Companies can fine-tune Timestral on their own data through Mistral Forge, the company's fine-tuning platform, to adapt it to specific industries 1.
Mistral says the model is meant to fit into existing monitoring systems and can work alongside its language models in multi-agent workflows, where an AI agent reacts to alerts from the forecasting system. The benchmark results have been submitted to the GIFT-Eval and TIME leaderboards for independent verification 1.
Key facts
- Timestral is Mistral's first family of time-series foundation models, available in four sizes from 5M to 209M parameters
- The 209M model ranks first on GIFT-Eval, TIME, and fev-bench, using only two-thirds the parameters of prior leaders
- Models run on standard CPUs — the 5M variant forecasts a day ahead in 10 milliseconds
- Timestral supports multivariate forecasting, categorical inputs, and known-future covariates
- Fine-tuning is available through Mistral Forge for domain-specific adaptation