# Mistral launches Timestral, a family of time‑series forecasting models

> The new models range from 5M to 209M parameters and top public benchmarks for multivariate, probabilistic forecasts.

Oossa · 2026-10-15 · https://oossa.com/en/mistral-launches-timestral-a-family-of-time-series-forecasting-models

Mistral announced the release of Timestral, its first set of time‑series foundation models. The biggest model has 209 million parameters and ranks first on the three main public forecasting benchmarks—GIFT‑Eval, TIME and fev‑bench—while using fewer parameters than earlier leaders. Four sizes are offered (5 M, 28 M, 123 M, 209 M) and they run on standard CPUs, so no GPU is needed for inference. The models support past‑only data, known‑future covariates and give probabilistic forecasts with 21 quantiles.

## The facts

- 209 M‑parameter model tops GIFT‑Eval, TIME and fev‑bench benchmarks
- Four model sizes are available: 5 M, 28 M, 123 M, 209 M

## Why it matters

Enterprises can improve sensor‑based planning and risk assessments without buying GPU hardware.

## Sources & references

1. [Introducing Timestral](https://mistral.ai/news/timestral/) – Mistral, 2026-10-15
2. [Quantile-Guided Diffusion for Probabilistic Time-Series Forecasting](https://arxiv.org/abs/2610.20739) – arXiv cs.LG, 2026-10-16
3. [New repository mistralai/timestral](https://github.com/mistralai/timestral) – Mistral, 2026-10-15

Last updated: 2026-10-16