Mistral releases Timestral, a time-series foundation model family
Mistral released Timestral, its first family of time-series foundation models, spanning 5M, 28M, 123M and 209M parameters with a shared architecture and training recipe. Mistral says its largest 209M-parameter model ranks first on the GIFT-Eval, TIME and fev-bench public benchmarks, beating larger models on key metrics, with results submitted to the GIFT-Eval and TIME leaderboards. The models produce probabilistic multivariate forecasts, accept past-only and known-future covariates and run without a GPU; Mistral cites roughly 10 ms for the 5M model and 140 ms for the 209M model forecasting one day ahead from three weeks of 15-minute history on an 8-thread server CPU. Code is available on GitHub under a modified MIT licence, with fine-tuning and domain adaptation offered through Mistral Forge and a technical report to follow.
Time-series forecasting is a common enterprise workload, so a small, CPU-deployable and fine-tunable open model family could lower the barrier to deploying forecasting in sensor, supply-chain and finance settings.