Timestral: Mistral release time-series foundation models
Points and comments are a snapshot, not live.
Mistral releases Timestral, a family of time-series foundation models up to 209M parameters.
Timestral delivers probabilistic, multivariate forecasts and supports past-only and known-future covariates. The largest model (209M parameters) ranks first on GIFT-Eval, TIME, and fev-bench, outperforming larger models. It uses decoupled horizon forecasting, covariate-aware attention, and synthetic-first data with two-stage training. Available in four sizes (5M, 28M, 123M, 209M), it runs efficiently on CPUs and offers zero-shot inference. Fine-tuning and domain adaptation are available via Mistral Forge.
What commenters are saying
Commenters compare Timestral favorably to existing open and proprietary models. A cofounder of The Forecasting Company notes their t0-gamma model (Apache 2 license) has a 42% winrate against Timestral on fevbench, while Timestral uses a restrictive modified MIT license. Some express skepticism about time-series foundation models' real-world utility, arguing specialized models often outperform. Others defend the approach, citing production use at quant funds and retailers. A user praises the speed and ease of use compared to building models from scratch.