Timestral: Mistral's time-series forecasting models, 5M to 209M

Mistral has released Timestral, its first family of time-series foundation models. They produce probabilistic, multivariate forecasts and are meant for data such as sensor telemetry, observability metrics, capacity planning, energy, supply chains and finance.

The family comes in four sizes: 5M, 28M, 123M and 209M parameters. Mistral says the largest model ranks first on GIFT-Eval, TIME and fev-bench among pre-trained, zero-shot models, using a single checkpoint without ensembling. It supports past-only and known-future covariates, including categorical ones, and predicts a full set of quantiles instead of one value.

Timestral needs no GPU at inference. On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model. Mistral says the 28M model beats Chronos-2, which is about four times larger, on GIFT-Eval and TIME. Fine-tuning for a specific domain is offered through Mistral Forge.

Timestral is a numerical model, not a language model. Mistral says it can run inside existing monitoring pipelines or be paired with its language models in agentic workflows, such as flagging a pump anomaly and then triggering a root cause analysis.