Mistral launches Timestral, its first family of time-series forecasting models

Mistral released Timestral on October 15, its first family of time-series foundation models. The four sizes range from 5 million to 209 million parameters, and all produce probabilistic, multivariate forecasts without task-specific training.

What Mistral claims

The company says the largest model ranks first among pretrained, zero-shot models on three public forecasting benchmarks: GIFT-Eval, TIME and fev-bench. It also says its 28M model beats Chronos-2 on GIFT-Eval and TIME despite being about four times smaller, and that the 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters. Mistral says it submitted the results to the GIFT-Eval and TIME leaderboards; these are the company’s own figures.

Built to run on CPUs

Timestral does not need a GPU at inference. Mistral reports about 10 ms for the 5M model and 140 ms for the 209M model on a standard 8-thread server CPU, for a one-day forecast from three weeks of 15-minute history. The models accept past-only and known-future covariates, such as a holiday calendar or a weather forecast, and can be fine-tuned through Mistral Forge.

Why it matters

Forecasting sensor, demand and financial data is a common enterprise need. Small models that run on ordinary servers lower the barrier to using foundation models there.