Mistral’s Timestral forecasts 21 possible futures—and the smallest model fits in 21 MB

Mistral has released Timestral, its first family of foundation models built for time series rather than language. The four models range from 5 million to 209 million parameters, and every one returns a probability distribution over the future instead of a single line pretending to be certain. For each future time step, the system produces 21 quantiles, from very low outcomes through the median to very high outcomes.[1][2]

That makes the smallest member unusually concrete. Timestral-5M is distributed as about 21 MB of weights and can run on a CPU. The largest model is about 837 MB. Mistral says all four sit on the accuracy-versus-size Pareto frontier in its launch comparisons, meaning that a larger alternative did not clearly dominate them on both dimensions at once.[2][1]

A forecast should show how wrong it might be

Suppose an electricity operator asks for tomorrow’s demand at 6 p.m. A conventional model might return 42 gigawatts. Timestral can instead return a sequence such as a 10th-percentile forecast of 38 GW, a median of 42 GW and a 90th-percentile forecast of 47 GW. Those numbers do not describe 21 separate universes. They mark points on one estimated distribution, giving the operator a way to plan for an ordinary outcome and a stressful one.

Known-future information has its own lane

Time-series forecasting often has information that is genuinely available in advance. A retailer knows the date of a promotion. A grid operator knows the calendar and may know a planned maintenance window. A hotel knows the booked-room count even though final occupancy is uncertain. Timestral separates past-only covariates from known-future covariates, so the model can use a planned event without allowing future target values to leak backwards into the prediction.[1][2]

The model is multivariate as well. Related series can be forecast jointly rather than one at a time. That matters when demand in several stores, temperatures at several sensors or prices across connected markets move together. The repository also supports missing values and categorical covariates, which makes the released software more than a bare research checkpoint.[2]

Synthetic data supplied most of the first lesson

Mistral says the majority of pre-training data was synthetic. The training process first generated many univariate processes with trends, seasons and noise, then constructed multivariate series by adding nonlinear relationships. A second stage mixed real and synthetic data and extended the context from 8,192 to 32,768 time steps. The design is intended to teach reusable temporal patterns before adaptation to a company’s own data.[1]

This is a sensible answer to a common forecasting problem: public time-series data is much scarcer and more fragmented than public text. It also creates a validation obligation. Synthetic diversity is not proof that a model understands one factory’s failure modes, one country’s power market or one retailer’s promotions. Backtesting must preserve the actual decision horizon and avoid letting later information enter the earlier training window.

The benchmark lead is a starting point, not the deployment decision

Mistral reports that Timestral-209M ranked first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench. The repository identifies 97 GIFT-Eval configurations, 98 TIME tasks and 100 fev-bench tasks. Those results make the release worth testing, especially because the models are compact, but the company also says a full technical report will follow.[1][2]

The practical comparison is therefore not “Does Timestral win a public table?” It is whether the smallest model that meets a particular accuracy target can update quickly enough, stay calibrated when conditions change and beat the organisation’s current seasonal, statistical or specialist forecasting baseline. A 21 MB model that can answer that question locally may prove more consequential than a much larger model that leads by a small average margin.

Report an error or suggest a correction