Mistral AI has released Timestral, its first family of time-series foundation models, in four sizes from a tiny 5M to a flagship 209M parameters. The models do zero-shot probabilistic forecasting — hand them historical data and they predict what comes next, with no training on your own data first — and Mistral says the 209M version ranks first among zero-shot forecasting models on three public benchmarks: GIFT-Eval, TIME and fev-bench.

The hook is efficiency. All four sizes share one architecture, and the entire family is small enough to run on a standard CPU — no GPU required. Mistral reports the 209M model produces a day-ahead forecast from three weeks of 15-minute history in around 140 milliseconds on an 8-thread server CPU, with the 5M model doing it in about 10 ms. That puts serious forecasting within reach of ordinary hardware, from analysts' laptops to edge devices.

Under the hood, Timestral outputs 21 quantiles per forecast step, giving a band of likely values rather than a single number — useful for sales planning, finance and operations, where knowing the range of outcomes matters more than a point estimate. It handles multivariate inputs, missing values (as NaN), and both past-only and known-future covariates like holidays, planned promotions or scheduled deploys. The largest model reads up to 32,768 past steps and forecasts up to 1,024 steps per pass.

One important caveat: the benchmark claims are Mistral's own — the technical report is still pending, and the margins over the previous leader (Paris rival The Forecasting Company's t0-gamma) are narrow. Still, two European teams now top the leaderboards in a field where Google's TimesFM and Amazon's Chronos were the names to beat.

On licensing, the weights ship under what Mistral calls a modified MIT license — with a notable modification: companies whose global monthly revenue exceeded $20 million in the preceding month are not authorized to use the weights. Mistral also shared a lightweight inference library and a public GitHub repository with the release.

Sources