Mistral launches Timestral time-series models
Original: Timestral: Mistral's time-series foundation models
Why This Matters
Zero-shot time-series forecasting at this scale addresses a high-value enterprise data gap.
Mistral released Timestral, a family of four time-series foundation models (5M–209M parameters). The 209M flagship ranks first on GIFT-Eval, TIME, and fev-bench benchmarks, beating larger models in zero-shot forecasting.
Mistral's first time-series foundation model family, Timestral, tops all three major public forecasting benchmarks — GIFT-Eval, TIME, and fev-bench — using just 209M parameters. That's roughly two-thirds the size of the previous state-of-the-art models it beat, from a single checkpoint with no ensembling.
The family spans four sizes: 5M, 28M, 123M, and 209M parameters, sitting on what Mistral describes as the Pareto front of accuracy versus model size. All models support probabilistic, multivariate forecasting, along with past-only and known-future covariates. They run efficiently on CPUs, making deployment lighter than GPU-heavy alternatives.
Classical tools like ARIMA learn from a single series in isolation. Deep learning models like PatchTST handle nonlinear patterns across related series but must be retrained per dataset. Timestral uses the pre-training paradigm — trained on a large, diverse corpus — to transfer learned patterns zero-shot to unseen domains, even where labeled data is scarce. Core techniques include synthetic pre-training, covariate-aware attention, and long-context learning. Fine-tuning for specific domains is available through Mistral Forge.
Target use cases span industrial sensor monitoring, equipment drift detection, capacity planning, financial flow forecasting, and supply chain management.