Mistral releases Timestral time-series forecasting models

It supports both univariate and multivariate forecasting, with predictions across 21 quantiles.[3] Mistral says it can produce forecasts for multiple related series and accept past-only or known-future covariates.[1] The model accepts up to 32,768 historical time steps and can forecast up to 1,024 steps in one forward pass.[3] These specifications give teams concrete limits to compare with their own history lengths and forecast horizons.

Mistral offers Timestral in 5M, 28M, 123M and 209M parameter versions.[1] The company says all four sit on the accuracy-versus-model-size Pareto front, and that the 123M version comes within 0.2% of the 209M model’s score at three-quarters of its latency.[1] Mistral says inference does not require a GPU.[1] On an eight-thread server CPU, it reports that forecasting one day ahead from three weeks of 15-minute history takes 10 milliseconds with Timestral-5M and 140 milliseconds with Timestral-209M.[1] It also says the 28M model is five times faster than the 209M.[1]

Mistral says its largest model ranks first on all three of the field’s main public benchmarks and outperforms the previous state of the art using two-thirds as many parameters.[1] It says those benchmark results use a single checkpoint, without ensembling.[1] Published scores for Timestral-209M include 0.658 on GIFT-Eval’s MASE metric, where lower is better, and 50.39 on fev-bench’s SQL skill score, where higher is better.[3] Separately, Superlinear reported that Timestral and TimesFM had 4.5% to 6.3% less forecast error than ETS on Walmart and FreshRetailNet grocery data, without training on those datasets.[6]

Mistral says most of Timestral’s training data was synthetic.[1] Its stated process first pretrains on synthetic data with an 8,192-step context, then post-trains on mixed real and synthetic data with a context four times longer.[1] Mistral says fine-tuning on proprietary data can specialize a model to a domain while preserving general performance.[1] The company published a GitHub repository for running Timestral on users’ own data.[1] The model weights are licensed under Modified MIT, while the public inference-code repository is licensed under Apache 2.0.[4][5]