October 19, 2026
Mistral Releases Timestral, Its First Time-Series AI Model Family
Mistral has released Timestral, its first family of time-series foundation models, designed to forecast data like sensor readings, energy demand, and financial flows. The family comes in four sizes, from 5M to 209M parameters, and the largest ranks first on all three major public forecasting benchmarks: GIFT-Eval, TIME, and fev-bench.
The models give probabilistic forecasts, meaning they show the range of likely outcomes and uncertainty instead of just one predicted number. They can combine many related data series at once, handle known future factors like production schedules or weather forecasts, and work on data types they were never trained on, which helps in fields where data is scarce.
A practical highlight is efficiency. The smallest model can forecast a day ahead from three weeks of 15-minute sensor history in 10 milliseconds on a standard CPU, with no GPU required. Timestral can be fine-tuned on private data through Mistral Forge, and the full pipeline can be deployed on a client's own hardware, including air-gapped environments.