Timestral: Mistral release time-series foundation models
Summary
Mistral unveils Timestral, a family of time-series foundation models with 5M–209M parameters that deliver state-of-the-art probabilistic, multivariate forecasts, including zero-shot inference. The models excel on CPU, support past and known-future covariates, and are designed for industrial and financial forecasting scenarios, with deployment options via Mistral Forge and on-premise or air-gapped environments. Benchmarks show Timestral leading on GIFT-Eval, TIME, and fev-bench, with a clear roadmap and available GitHub resources.