AI Research
Oct 19, 2026
Mistral launches Timestral, a new family of time-series foundation models
Oct 19, 2026
AI Summary
Mistral has introduced Timestral, its first series of time-series foundation models designed for probabilistic and multivariate forecasting. The models, ranging from 5M to 209M parameters, demonstrate superior performance on key benchmarks while being efficient for various applications, including industrial and financial forecasting.
- Timestral is Mistral's first family of time-series foundation models, optimized for probabilistic and multivariate forecasts.
- The models come in four sizes: 5M, 28M, 123M, and 209M parameters, with the largest model outperforming competitors on major benchmarks with fewer parameters.
- Timestral supports both past-only and known-future covariates, allowing for improved forecasting accuracy in various domains.
- The models are designed to run efficiently on standard CPUs, making them accessible for organizations without extensive GPU resources.
- Key features include multivariate forecasting, probabilistic outputs with quantiles, and the ability to handle categorical and known-future covariates directly.
- Timestral has been evaluated on public benchmarks such as GIFT-Eval, TIME, and fev-bench, achieving top rankings across multiple metrics.
- The model architecture allows for zero-shot forecasting, enabling it to make predictions on unseen data.
- Mistral Forge offers options for fine-tuning the models on proprietary data and adapting them to specific domains.
- The models can be deployed in various environments, including on-premises and air-gapped systems, ensuring flexibility for different organizational needs.
time-seriesfoundation modelsmistralai researchmachine learning