Mistral launches Timestral, a time-series model family that tops three benchmarks

Four sizes, probabilistic forecasts

Timestral comes in four sizes: 5M, 28M, 123M and 209M parameters. Mistral says all four share the same architecture and training recipe. The larger models improve forecast performance, while the smaller ones are faster and cheaper.

The models deliver probabilistic, multivariate forecasts and support past-only and known-future covariates. Timestral predicts a full set of quantiles, so teams can see how wide the uncertainty is and set alert thresholds that reflect it. It also accepts categorical covariates, such as the day of the week or a machine's state, directly as categorical inputs instead of encoding them as numbers.

Mistral positions the models as time series foundation models. They are trained on a large and diverse corpus and forecast zero-shot on data they have never seen, even in domains where data is scarce. This contrasts with classical models such as ARIMA, which are usually fitted to a single series, and with deep learning models such as PatchTST, which are typically trained from scratch on each dataset.

Benchmark and efficiency claims

Mistral evaluated Timestral on GIFT-Eval, TIME and fev-bench. Among pre-trained, zero-shot models, it says the model ranks first across all five metrics, using a single checkpoint without ensembling. Results have been submitted to the GIFT-Eval and TIME leaderboards.

The company says every size sits on the Pareto front of accuracy versus model size, meaning that for each size no public model forecasts better. Two comparisons stand out:

• The 28M model beats Chronos-2, which is about four times larger, on GIFT-Eval and TIME.
• The 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters.

Timestral does not need a GPU at inference, according to Mistral. 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.

How it was built

Mistral names four technical ideas and says a technical report will describe them in detail.

• Decoupled Horizon Forecasting. Causal decoders trained with contiguous patch masking, as in Toto 2.0 and TimesFM-3, mask out the values to be forecast. Mistral's approach keeps the full history in the sequence and uses learned forecast tokens that read the history before them, so every observed value serves as both target and context.
• Covariate-aware attention. Each input series is assigned a role: target variate, past-only covariate or known-future covariate. Information flows only from covariates to targets, so a holiday calendar or weather forecast can be used without leaking a target's own future.
• Synthetic-first data. The majority of the training data was synthetic, built from univariate processes with trends, seasonality and diverse noise, then combined into multivariate series with nonlinear dependencies.
• Two-stage training. The model is first pre-trained on purely synthetic data with context size 8192, then post-trained on a mix of real and synthetic data with a 4x longer context. The objective is a pinball loss over 21 quantiles.

Industrial use and adaptation

Mistral uses an industrial plant as its main example. Pump vibration, temperature and pressure do not move independently, and Timestral models related series jointly. Known drivers such as production schedules, planned shutdowns and weather forecasts can be fed in as known-future covariates, so a scheduled load increase is expected rather than flagged as a surprise.

The company describes Timestral as a numerical model that can run inside existing monitoring pipelines or be paired with its language models in agentic workflows. In its sketch, Timestral detects an anomaly on a pump, an agent compares the signals with past sensor data to find similar failures, and sub-agents run a root cause analysis using maintenance records, shift logs and manuals. The plant manager then receives a suggested diagnosis with evidence and next steps rather than just an alert.

For organisations whose data differs from public benchmarks, Timestral is available through Mistral Forge, the training pipeline Mistral uses for its language models. It supports zero-shot inference out of the box, architectural extensions such as domain-specific covariates, and fine-tuning on proprietary data.

What it means for users

For teams in energy, finance, supply chains and industrial maintenance, the release offers forecasting models that, per Mistral, run on CPUs and range from 5M to 209M parameters. That lets users trade accuracy for speed and cost. The benchmark rankings are the company's own claims, and the technical report promised by Mistral should show more detail on how the results were reached.