Your team forecasts tomorrow’s hourly support volume for staffing. On October 15, 2026, Mistral introduced Timestral in four sizes from 5 to 209 million parameters. It forecasts several series and returns values at different probability levels. It accepts known-future features such as holidays and past-only data such as observed volume. The official repository supports CPU and GPU. Code uses Apache 2.0 and weights use a Modified MIT license.
Test with your own data
On Linux, pin revisions of code, dependencies and weights. Create eight weeks of synthetic hourly volume, with holidays and an unusual week. Replay each morning in the last two weeks using only what existed at that time. Compare all four sizes with and without calendar inputs against the same day one week earlier. Fix CPU thread count. Record mean absolute error, runtime and peak memory. Measure the share of observations within the 10th-to-90th percentile interval.
Making a decision
Mistral’s benchmark results apply to its test conditions. Choose the smallest size meeting your error and resource targets on important days. Our TimesFM‑3 article explains multivariate forecasting. This trial focuses on choosing a CPU model size.