Timestral is Mistral AI's first set of foundation models for time series, released on 15 October 2026, which forecast future values of data recorded over time.
In plain words
A time series is a list of measurements taken over time, such as sensor readings. Timestral learned patterns from a large mix of series, so it can forecast data it has never seen without extra training. It comes in four sizes, from 5 million to 209 million parameters, and runs on ordinary CPUs. It gives a range of likely values rather than one number, and can use extra inputs like which weekday it is. Mistral says the largest ranks first on three main public benchmarks.
A quick example
Imagine a bakery that logs how many loaves it sells every hour. It feeds three weeks of that history to a forecasting model, along with which days are holidays, and gets a likely range for tomorrow's sales.
Where you'll meet it
- Planning capacity, energy use and supply chains from past measurements.
- Spotting equipment drift early by forecasting factory sensor readings.
Related
Sources
- DocsIntroducing Timestral, Mistral AI · read 15 Oct 2026
- DocsTimestral-209M model card, Mistral AI (Hugging Face) · read 15 Oct 2026
Checked by WhatIsAI against the sources above · 15 Oct 2026. A full five-minute explainer with a diagram may follow. Spotted an error? Report it.