Mistral releases Timestral time-series models that top forecasting benchmarks

Mistral enters time-series forecasting

Mistral has released Timestral, its first family of time-series foundation models. According to the company's announcement, the largest model, at 209M parameters, ranks first on all three of the field's main public benchmarks — GIFT-Eval, TIME and fev-bench — among pre-trained, zero-shot models, with results from a single checkpoint and no ensembling. On GIFT-Eval it places in the top ten of 83 comparable pretrained entries, on TIME in the top ten of 35 leaderboard entries, and on fev-bench in the top ten of 37. Mistral says results have been submitted to the GIFT-Eval and TIME leaderboards.

The family comprises four sizes — 5M, 28M, 123M and 209M parameters — all sharing the same architecture and training recipe. Mistral claims that at each size no public model forecasts better, putting every variant on the Pareto front of accuracy versus size. The 209M model reportedly beats the previous state of the art using only two-thirds of the parameters, outperforming TimesFM-3 and Falcon-2.0, while the 28M model beats Chronos-2, roughly four times larger, on GIFT-Eval and TIME.

Probabilistic, multivariate forecasting with covariates

Timestral produces probabilistic, multivariate forecasts. Instead of a single predicted value it outputs a full set of quantiles, letting teams see how wide the uncertainty band is, set thresholds accordingly, and plan maintenance around risk rather than a point estimate. Because related series are modelled jointly, a rise in a pump's temperature alongside its vibration can inform the forecast instead of each sensor being handled in isolation.

The models accept both past-only and known-future covariates. Known-future drivers such as production schedules, planned shutdowns or weather forecasts are visible across the whole window while targets stay causal, and information flows only from covariates to targets — so a holiday calendar improves a forecast without the target's own future leaking into it. Categorical covariates, such as day of week or a machine's state (running, idle, under maintenance), are taken as categorical inputs directly rather than encoded as numbers.

Designed to run on CPUs

No GPU is needed at inference. Mistral reports that on a standard server CPU with 8 threads, producing a one-day-ahead forecast from three weeks of 15-minute data takes 10 ms with the 5M model and 140 ms with the 209M model. Smaller sizes trade some accuracy for speed: 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.

How it was trained

Mistral attributes the results to four technical contributions, to be detailed in a technical report:

• Decoupled Horizon Forecasting (DHF). Causal decoders trained with contiguous patch masking, as in Toto 2.0 and TimesFM-3, lose the masked values as context for later forecasts. DHF keeps the full history in the sequence and lets learned forecast tokens read the history before them, so every observed value 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. Each input series is assigned a role: target variate, past-only covariate or known-future covariate.
• Synthetic-first data. The majority of the training corpus was synthetic — univariate processes with trends, seasonality and diverse noise patterns, then combined by a multivariate synthesizer that introduced nonlinear dependencies between them.
• Two-stage training. The model is pre-trained at scale on purely synthetic data with a context size of 8192, then post-trained on a mix of real and synthetic data with a 4x longer context, optimising a pinball loss over 21 quantiles, a discrete approximation of CRPS.

From forecasts to workflows

Mistral frames Timestral as a numerical model that can run inside existing monitoring pipelines or be paired with its language models in agentic workflows. In a sketched industrial scenario, Timestral watches sensor streams continuously and flags anomalies; an agent embeds the flagged signals, retrieves similar historical failures, and spawns sub-agents for root-cause analysis drawing on maintenance records, shift logs, equipment manuals and readings from related equipment — ending with a suggested diagnosis and plan rather than a bare alert. Fine-tuning for domain adaptation is available through Mistral Forge.

Why it matters

Time series is among the most common data types in enterprise — from sensor telemetry to observability metrics — and forecasts built on it drive capacity planning, energy, supply chains, risk and maintenance. Classical tools such as ARIMA fit a single series and learn only from that series' history; deep-learning models such as PatchTST learn nonlinear patterns across related series but are typically trained from scratch per dataset, so what they learn does not carry over to a new domain. Foundation models apply the pre-training recipe to time series, forecasting zero-shot on data they have never seen, even where data is scarce. If the claims hold up on the public leaderboards, a sub-1B-parameter family that runs on CPUs and beats much larger competitors makes competent forecasting a cheap default for observability and monitoring stacks rather than a specialist project.