Timestral forecasts related time series with known future inputs
Mistral’s Timestral models forecast multiple related numerical series together and return uncertainty ranges instead of only one predicted value. They can use historical inputs and information known in advance, such as a holiday calendar, while preventing future target values from entering the forecast. Four sizes range from 5 million to 209 million parameters and support CPU inference. Production usefulness still depends on testing against the actual data distribution.
Mistral AI · Oct 15, 2026
Why it matters
Capacity and monitoring forecasts often need scheduled future events and uncertainty, not just extrapolation from a single metric.
More context
The public inference repository requires Python 3.10 or newer and PyTorch 2.4 or newer. Code uses Apache 2.0; weights use a separate Modified MIT license. Try the smallest model on a held-out time window and compare it with a simple forecasting baseline before adopting it.
Keep in mind
Mistral’s benchmark rankings are vendor-reported, and the full technical report is forthcoming. Forecast quality may change with domain shifts, missing data and the availability of reliable future inputs.
Original read