AI news · October 16, 2026
Mistral launches Timestral forecasting models from 5M to 209M parameters
Mistral released Timestral models from 5M to 209M parameters, with a claimed 140ms one-day forecast on eight CPU threads, so time-series features can start without a GPU bill.
- largest model
- 209M
- one-day forecast
- 140ms
- forecast quantiles
- 21
Mistral released Timestral on October 15 as a family of probabilistic multivariate forecasting models sized at 5M, 28M, 123M and 209M parameters. The company reports that the 209M model produces a one-day forecast from three weeks of 15-minute data in 140 milliseconds on eight CPU threads. It supports 21 quantiles, a 32,768-token context and on-premise, VPC or air-gapped use. Mistral reports leading results on GIFT-Eval, TIME and fev, and says the 28M model beats Chronos-2, which is about four times larger.
The practical change is lower deployment cost for demand, sensor and operations forecasts. Timestral is not a general language model, and the benchmark claims still need a check on the series a business owns. The open repository gives a faster start than building a forecasting service around a large chat model, while CPU support may suit edge or scheduled jobs. Accuracy, calibration and retraining cost remain the commercial tests.
What you can do with it
Try the Mistral repository on one historical series with a time-based holdout. Compare Timestral 28M or 209M with a simple seasonal baseline and Chronos-2; record error by horizon, CPU time, memory and the cost of refreshing the model before adding it to a dashboard or alert.
Our take
This is a strong practical release for products with repeated numeric data. A fast CPU model can beat a chat model on cost and consistency, but the headline scores do not prove value on your series. A small backtest is enough to decide whether to continue.
Source: Mistral AI — Made With Models writes the brief; the reporting is theirs.