Mistral Releases Timestral Time-Series Models for Multivariate Forecasting and Known Future Covariates

Mistral released Timestral, its first time-series foundation model family, on October 15. It is available in four sizes—5M, 28M, 123M, and 209M parameters—and is designed for numerical forecasting tasks involving sensors, demand, and capacity. The models support zero-shot inference, allowing teams to test new data with existing weights before deciding whether domain-specific fine-tuning is needed. Official announcement

The largest version uses a 24-layer Transformer with temporal and variable attention. It groups every 32 time steps into a patch and encodes missing-value indicators alongside them. The model accepts up to 32,768 historical time steps, predicts 1,024 steps in a single forward pass, and outputs 21 quantiles. Longer forecasts are generated in chunks, feeding the previous chunk’s median predictions back into the context; error accumulation in long-range forecasts should therefore be tested separately. Model card

The technical focus is on retaining historical information and using external signals. Decoupled Horizon Forecasting lets learnable forecast tokens read from the history, avoiding confusion between forecast targets and genuinely missing values. Attention also distinguishes targets, historical-only covariates, and covariates known in advance, allowing schedules or holiday calendars to inform forecasts while limiting information leakage from future target values. Training began with purely synthetic data, then incorporated real data as the context length was extended. Technical explanation

The official model card reports a GIFT-Eval MASE of 0.658 for the 209M version and 0.659 for the 123M version. Their corresponding TIME scores are 0.626 and 0.628; lower is better for all these metrics. The smaller versions are already close to the largest model on these two benchmarks, making them worth including in deployment cost comparisons. However, these results are from the publisher, and a full technical report has not yet been released. Benchmark table

Engineering teams can run the models on CPUs or GPUs using PyTorch. The weights are licensed under the Modified MIT License. For monitoring systems, quantile outputs can inform alert thresholds; before deployment, teams should use time-based backtesting to examine quantile coverage, data drift, and whether future covariates will actually be available at forecast time. Model size and public leaderboard results alone are not enough to predict accuracy or latency on a specific production line. Deployment and license information