Mistral has released Timestral, its first family of time-series foundation models. The flagship model, with 209M parameters, ranks first on the GIFT-Eval, TIME and fev-bench benchmarks, outperforming the previous state of the art while using only two-thirds of its parameters. Timestral produces probabilistic, multivariate forecasts and supports both past-only and known-future covariates.
The family comes in four sizes, from 5M to 209M parameters, and each size sits on the Pareto front of forecasting accuracy versus model size. The 28M model beats Chronos-2, which is about four times larger, on GIFT-Eval and TIME. Timestral also requires no GPU at inference. On a standard server CPU, forecasting one day ahead takes 10 ms with the 5M model and 140 ms with the 209M model.
Time-series data is common in enterprise settings, from sensor telemetry to observability metrics, and drives decisions in capacity planning, energy, supply chains, risk and maintenance. Classical models like ARIMA are usually fitted to a single series, while deep learning models such as PatchTST are typically trained from scratch on each dataset. Foundation models instead learn patterns from a large, diverse corpus that transfer across series, allowing zero-shot forecasting even where data is scarce.
The model combines synthetic pre-training, covariate-aware attention and long-context learning, and can be fine-tuned for specific domains through Mistral Forge. Results have been submitted to the GIFT-Eval and TIME leaderboards.
Source: Hacker News · Summarized by HeadlinesBriefing