Mistral released Timestral on October 15, 2026, saying it delivers state-of-the-art probabilistic, multivariate forecasts with just 209M parameters. It is the company's first time-series foundation model update since its initial model launch.
Mistral reported its largest model ranks first on all three of the field's main public benchmarks, GIFT-Eval, TIME, and fev-bench, measured on forecasting accuracy. That compares with previous state-of-the-art models that required more parameters.
Timestral is built on synthetic pre-training, covariate-aware attention, and long-context learning and targets forecasting across diverse domains. Availability begins with four sizes, from 5M to 209M parameters, initially for industrial and financial use cases.
"Today, we’re releasing Timestral, our first family of time-series foundation models," said Mistral. "Our largest model, at just 209M parameters, ranks first on all three of the field's main public benchmarks."
The announcement follows Mistral's previous model releases and its focus on time-series forecasting. Mistral emphasized the significance of Timestral in enabling accurate forecasts for domains where data is scarce.
Mistral did not say how it will scale Timestral further, and it raised the open question of how the model will perform on even more complex time-series data. The company said it will continue to refine the model and expand its applications.
Source: mistral
