Mistral has released Timestral, its first family of time-series foundation models designed for forecasting tasks across industrial and financial applications. The largest model, containing 209M parameters, ranks first on all three major public benchmarks—GIFT-Eval, TIME, and fev-bench—while using approximately two-thirds of the parameters of previous state-of-the-art models.
The Timestral family consists of four models optimized for different use cases: 5M, 28M, 123M, and 209M parameters. According to Mistral, each model sits on the Pareto front of forecasting accuracy versus model size, meaning no public model of similar size performs better. The 28M model outperforms Chronos-2, which is roughly four times larger, on GIFT-Eval and TIME benchmarks.
Unlike classical forecasting approaches such as ARIMA and linear regression, which are typically fitted to individual series, Timestral uses a foundation-model approach. The models are trained on diverse data and can transfer learned patterns across different time series, including zero-shot forecasting on unseen domains. This addresses a limitation of deep learning models like PatchTST, which must typically be trained from scratch on each new dataset.
Timestral delivers probabilistic, multivariate forecasts and supports both past-only and known-future covariates. The models can handle categorical covariates directly—such as day of the week or machine state—without imposing artificial ordering. They produce quantile predictions rather than single-point estimates, enabling teams to understand uncertainty ranges and set alert thresholds accordingly.
The models are computationally efficient, running on standard server CPUs without requiring GPUs. On an 8-threaded CPU, the 5M model generates one-day-ahead forecasts in 10 milliseconds from three weeks of 15-minute data, while the 209M model takes 140 milliseconds. The 28M model is five times faster than the 209M version with minimal accuracy loss.
Mistral’s technical approach includes several innovations: decoupled horizon forecasting that preserves full history as context, covariate-aware attention that prevents information leakage from future values, and synthetic-first pre-training followed by post-training on mixed real and synthetic data. Results have been submitted to the GIFT-Eval and TIME leaderboards.
Key facts
- Timestral’s 209M-parameter largest model ranks first on GIFT-Eval, TIME, and fev-bench benchmarks
- The family includes four models: 5M, 28M, 123M, and 209M parameters
- The 28M model outperforms the larger Chronos-2 model on major benchmarks
- Models run on standard CPUs without GPU requirements; 5M model forecasts in 10ms, 209M in 140ms
- Timestral supports multivariate forecasting, probabilistic predictions, and both categorical and continuous covariates
- The models use transfer learning, enabling zero-shot forecasting on unseen domains