Mistral Releases Timestral, Its First Family of Time-Series Foundation Models

Timestral Sets New Benchmark for Forecasting

Mistral has released Timestral, its first family of time-series foundation models, delivering probabilistic, multivariate forecasts. The largest model has 209M parameters and ranks first on all three main public benchmarks, GIFT-Eval, TIME and fev-bench, outperforming the previous state of the art with only about two-thirds of the model parameters. The results come from a single checkpoint, without ensembling, and the model supports past-only and known-future covariates.

A Family Built for Every Specification

Timestral is available in four sizes, 5M, 28M, 123M and 209M parameters, all sharing the same architecture and training recipe. Each size sits on the Pareto front of forecasting accuracy versus model size. The 28M model beats Chronos-2, which is about four times larger, on GIFT-Eval and TIME, while the 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters.

None of the models needs a GPU at inference. On a standard server CPU with 8 threads, 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. The 28M model is five times faster than the 209M model, and the 123M model comes within 0.2% of the 209M score at three quarters of its latency.

Technical Contributions

Timestral combines decoupled horizon forecasting, which keeps the full history in the sequence while learned forecast tokens read it, and covariate-aware attention that assigns each input series a role such as target variate, past-only covariate or known-future covariate. Information flows only from covariates to targets, so a target's own future never leaks into its forecast.

Training follows a synthetic-first approach. The model is pre-trained at scale on synthetic univariate and multivariate series with a context size of 8192, then post-trained on a mix of real and synthetic data with a 4x longer context, optimising a pinball loss over 21 quantiles as a discrete approximation of CRPS.

Availability and Adaptation

A GitHub repository is available to run Timestral on proprietary data, and a technical report with the full results will follow. For production adaptation, Timestral is offered through Mistral Forge, the same training pipeline used for Mistral's language models, supporting zero-shot inference out of the box, architectural extensions for domain-specific covariates, fine-tuning on proprietary data, and full control over the training mix including synthetic data generation. Deployment is possible on-prem, in a VPC or in air-gapped environments, under a modified MIT licence.

Sources: mistral.ai