Mistral's 209M-parameter Timestral tops all three major time-series benchmarks

Timestral: Mistral release time-series foundation models

Mistral enters forecasting with Timestral, a family of time-series foundation models spanning 5M to 209M parameters. Its largest model ranks first on GIFT-Eval, TIME, and fev-bench among zero-shot pre-trained models, beating larger rivals like TimesFM-3 and Chronos-2 with two-thirds the parameters. It handles multivariate, probabilistic forecasts with past and known-future covariates, runs on plain CPUs, and is largely trained on synthetic data. Fine-tuning is available through Mistral Forge, with weights on GitHub.