Mistral.ai has released Timestral, a new family of time-series foundation models designed for enterprise forecasting. The largest model, with 209 million parameters, ranks first on the main public benchmarks GIFT-Eval, TIME, and fev-bench, outperforming previous state-of-the-art models while using only two-thirds of their parameters. Timestral supports probabilistic, multivariate forecasts and handles past-only and known-future covariates, according to mistral.ai.
Timestral models deliver leading accuracy relative to model size, placing on the Pareto front of forecasting performance. The models have been submitted to the GIFT-Eval and TIME leaderboards for validation. Time-series data is critical across industries, from sensor telemetry in industrial plants to financial flows. Mistral highlights applications such as capacity planning, equipment maintenance, and risk management where Timestral’s forecasts can improve decision-making.
Traditional forecasting methods like ARIMA and linear regression typically analyze single series independently, while recent deep learning models like PatchTST learn nonlinear patterns across multiple related series. Timestral advances this by providing multivariate, probabilistic forecasts with fewer parameters, enabling enterprises to efficiently forecast complex time-series data across thousands of assets or accounts. This positions Mistral’s models as competitive tools in the growing market for AI-driven forecasting solutions.
Timestral’s release marks a notable development in time-series AI, with its benchmark submissions publicly available on GIFT-Eval and TIME leaderboards. Enterprises leveraging these models can expect enhanced forecasting capabilities for operational and financial planning, supported by Mistral’s latest research and engineering efforts.