Mistral released Timestral on October 15, 2026, its first family of time-series foundation models. There are four sizes, 5 million, 28 million, 123 million and 209 million parameters, sharing one architecture and training recipe, under a Modified MIT license, with a Hugging Face collection and a GitHub repository for running them on your own data.
The models produce probabilistic, multivariate forecasts with 21 quantiles, accept past-only and known-future covariates, forecast zero-shot, handle a context of 32,768 time steps and run on CPU without a GPU. On an 8-thread server CPU, Mistral says 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.
Mistral says the 209M model ranks first on the GIFT-Eval, TIME and fev-bench benchmarks, beating the previous state of the art with about two-thirds of its parameters, and ranks first on all five metrics among pretrained zero-shot models from a single checkpoint without ensembling. It says the 28M model beats Chronos-2, about four times larger, on GIFT-Eval and TIME. Fine-tuning is offered through Mistral Forge, and a technical report is to follow.
Why it matters: a general-purpose LLM lab entering time-series forecasting, with models small enough to run on CPU, extends foundation-model methods to operational forecasting such as energy, demand and sensor data, a market previously led by Google’s TimesFM and Amazon’s Chronos.
What it does not show: the leaderboard claims are Mistral’s own, made before its technical report was published, and the announcement does not explicitly describe the weights as open, though it lists a license and links to downloadable models.