Mistral releases Timestral time-series forecasting models
Four sizes from 5M to 209M parameters, trained mostly on synthetic data and run on CPUs; Mistral says the largest ranks first on three public forecasting benchmarks.
- Models & capabilities
- Open weights & ecosystem
- Minor
Mistral AI released Timestral, which it called its first family of time-series foundation models: pretrained forecasting models intended to predict sensor, financial or demand data on series they have not seen before. It comes a week and a half after Mistral Large 4 and is a different kind of product, a numerical model rather than a language model.
The family has four sizes, at 5M, 28M, 123M and 209M parameters, with a context of 32,768 time steps. Mistral said the models produce probabilistic forecasts as 21 quantiles, handle multivariate targets, and accept both past-only covariates and known-future ones such as production schedules or weather forecasts. Most of the training data was synthetic, generated as univariate processes with trend, seasonality and noise and then combined into multivariate series; a second stage used a mix of real and synthetic data.
On benchmarks, Mistral said the 209M model ranks first among pre-trained, zero-shot models on GIFT-Eval, TIME and fev-bench, “with only two-thirds of the model parameters” of the previous state of the art. It said the 28M model beats Chronos-2, which it described as about four times larger, on GIFT-Eval and TIME. These are Mistral’s own figures: it said results had been submitted to the GIFT-Eval and TIME leaderboards, and the benchmark scores on the Hugging Face model card are marked as not verified. A technical report was promised.
Mistral stressed that no GPU is needed. On an 8-thread server CPU, it said, forecasting one day ahead from three weeks of 15-minute history takes 17 ms with the 5M model and 113 ms with the 209M model. The weights are on Hugging Face under what the announcement calls a modified MIT licence, with code on GitHub. Mistral also offers the models for fine-tuning on a customer’s own data through its Forge platform, including on-premises and air-gapped deployments.
The release extends Mistral from language and multimodal models into forecasting for industrial and financial customers, the use cases it chose to illustrate the models.