Mistral releases Timestral: a 209M-parameter model that tops three benchmarks

Mistral AI has released Timestral, its first line of foundation models for time-series forecasting. The flagship version has just 209M parameters, yet according to Mistral's results it ranks first on the three main public benchmarks: GIFT-Eval, TIME and fev-bench. The previous record holder was about a third larger. Among pretrained zero-shot models, Timestral leads on all five metrics, and the results come from a single checkpoint, without ensembling.

Why time series need foundation models

Every enterprise has time series: sensor telemetry, observability metrics, financial flows, inventory levels. Capacity planning, energy, supply chains and equipment maintenance are built on them. Classical ARIMA and linear regression train on the history of a single series and know nothing about its neighbors. Neural networks such as PatchTST capture nonlinear patterns across related series, but they are usually trained from scratch for each dataset, and what they learn does not transfer to new domains.

Timestral takes an approach familiar from language models: pretraining on a large, heterogeneous corpus, after which the model forecasts data it has not seen before. That helps where history is scarce. Mistral names three ingredients in the recipe: synthetic pretraining, covariate-aware attention and long-context training.

Sizes, speed and comparison with competitors

The family consists of four models that share one architecture and one training recipe. According to Mistral, each of them sits on the Pareto frontier of accuracy versus size: at its scale, no public model is better.

• Sizes: 5M, 28M, 123M and 209M parameters.
• The 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters.
• The 28M model outperforms Chronos-2, which is roughly four times larger, on GIFT-Eval and TIME.
• The 123M model trails the 209M by only 0.2% in score, but runs at three quarters of the latency.
• The 28M model is five times faster than the flagship.
• Sample sizes in the tables: 83 comparable entries in GIFT-Eval, 35 in TIME, 37 in fev-bench.

Inference needs no GPU. On a standard server CPU with 8 threads, a forecast one day ahead from three weeks of 15-minute history takes 10 ms for the 5M model and 140 ms for the 209M model. The results have been submitted to the GIFT-Eval and TIME leaderboards.

What it can do in practice

Mistral demonstrates the model's capabilities with an example of an industrial plant with thousands of sensors. The model forecasts related series jointly: a pump's vibration, temperature and pressure depend on each other, and Timestral uses those links instead of treating each sensor separately. The forecast is probabilistic, meaning the model outputs a set of quantiles. A team can see the width of the uncertainty, set alarm thresholds based on it and plan maintenance with risk in mind, rather than relying on a single number. The model accepts both covariates known only for the past and those known in advance, such as a production schedule. Categorical covariates are also supported.

Who should use it, and how

For teams whose forecasting lives in a collection of ARIMA scripts or homegrown neural networks, Timestral offers a ready-made baseline with no training: feed it the history and get quantiles back. Because it has no GPU requirements, the model can run next to the data on ordinary servers, and with thousands of series you can choose between the 28M for speed and the 209M for accuracy. For specific domains, Mistral offers fine-tuning through Forge. Before deploying, run the model on your own historical data and compare it with your current solution: public benchmarks show the general level, but they do not guarantee results on particular sensors or financial series.