On October 15, Mistral introduced Timestral in four sizes from 5 million to 209 million parameters. It is not a chatbot: it produces probabilistic forecasts from historical time series, using related variables and circumstances known in advance, such as a work calendar.

The developer released code and models and reports CPU operation and strong public-benchmark results. We ran no independent measurement; a full technical report is still to follow. Code uses Apache 2.0, while weights have a separate Modified MIT license with commercial conditions; the two must not be confused.

Imagine a small workshop forecasting tomorrow's electricity demand from measurements and planned shifts. Instead of a single seemingly precise number, it could examine a range of possible values. This might flag unusual behavior earlier, but does not itself explain the cause or establish that a machine will fail.

Requirements include clean timestamps, correct handling of missing values and testing strictly forward from training history. A future covariate must truly be known when the forecast is issued. Compare against a simple seasonal baseline and monitor false alarms and changes in operations.

Optimistically, 2–6 weeks of good preparation and backtesting could show whether a non-intervening pilot is worthwhile. Forecasting rare failures will often require much longer observation. This is an editorial scenario, not a promise of savings or permission to put a model in charge of equipment.