Mistral's 209M time-series model tops three public forecasting benchmarks, and it runs a one-day forecast on a CPU in 140 ms

Forecasting gains a pretrained foundation-model family that can be applied directly, so sensor, energy and supply-chain teams no longer have to train a model from scratch for every dataset. Classical ARIMA and linear regression are usually fitted to a single series and learn only from that series' history; deep models such as PatchTST learn nonlinear patterns across related series but are still trained from scratch per dataset, so what they learn does not carry to new domains. Across GIFT-Eval, TIME and fev-bench, Timestral ranks first on all five metrics among pretrained zero-shot models, from a single checkpoint without ensembling; the comparisons cover 83, 35 and 37 entries respectively.

The model forecasts related series jointly and takes categorical and known-future covariates directly, returning quantiles rather than a single value. Forecasting each sensor in isolation discards the dependencies among vibration, temperature and pressure, and a point forecast cannot express how wide the uncertainty is. It outputs 21 quantiles as a probabilistic forecast; covariates enter attention under three roles, target variate, past-only covariate and known-future covariate, with information flowing only from covariates to targets so a target's own future does not leak into its forecast.

Most training data is synthetic: the model is pretrained at scale on purely synthetic series with context 8192, then post-trained on a mix of real and synthetic data with a four-times-longer context. Real industrial and financial series are thinly covered in public corpora, so pretraining on real data alone would struggle to span diverse trends, seasonality and noise patterns. The synthetic stage first generates univariate processes with coherent temporal dynamics, then a multivariate synthesizer introduces nonlinear dependencies between them; the objective is a pinball loss over 21 quantiles, a discrete approximation of CRPS.

The demonstrated scope is pretrained zero-shot forecasting: first place among pretrained zero-shot entries on the public GIFT-Eval, TIME and fev-bench benchmarks, from a single checkpoint without ensembling. The four sizes are 5M, 28M, 123M and 209M parameters, with a 32768-step context, 21-quantile output, a modified MIT licence, and on-prem, VPC or air-gapped deployment. The latency case is a one-day-ahead forecast from three weeks of 15-minute history on an 8-thread standard server CPU. The releasing organization says a GitHub repository is published for running Timestral on your own data, with a full technical report to follow.

Public benchmarks measure general performance, while production data can differ significantly in behavior and structure and shifts as markets change regime, equipment degrades and demand patterns move; the releasing organization treats continued adaptation and retraining as the lasting value. Synthetic data makes up the majority of training data, and the real-to-synthetic mix used in post-training is not given. The accuracy-latency trade-offs are stated for specific cases, with the 28M model five times faster than 209M and the 123M model within 0.2% of the 209M score at three quarters of its latency, and these numbers apply to the described CPU case rather than every deployment.