Time-series forecasting just got a serious open-source contender. On October 15, Mistral released Timestral, a family of four foundation models purpose-built for time-series prediction — ranging from a 5-million-parameter model to a 209-million-parameter flagship.
The headline result: the 28M-parameter Timestral outperforms Chronos-2 (a competing model with roughly four times as many parameters) on two of the field's main benchmarks, GIFT-Eval and TIME. The top 209M model takes first place across all three major public benchmarks: GIFT-Eval, TIME, and fev-bench.
What it can do
Timestral supports:
- Univariate and multivariate forecasting — single series or multiple correlated signals
- Probabilistic predictions across 21 quantiles — so you get uncertainty estimates, not just point forecasts
- Up to 32,768 historical time steps as context, forecasting up to 1,024 steps ahead in a single pass
- Known-future covariates — you can tell the model what you already know about the future (planned promotions, holidays, scheduled outages)
And critically: Timestral supports CPU-based inference. You don't need a GPU to run it.
Four models, one decision
| Model | Parameters |
|---|---|
| Timestral 1.0 5M | 5 million |
| Timestral 1.0 28M | 28 million |
| Timestral 1.0 123M | 123 million |
| Timestral 1.0 209M | 209 million |
The 5M model is released under a Modified MIT license. The larger models are available as public previews through Mistral's API and on Hugging Face at mistralai/Timestral-209M.
Why it matters
Time-series forecasting is everywhere — demand planning, anomaly detection, financial modeling, energy grids, weather prediction. The problem has traditionally required either expensive purpose-built models or adapting general LLMs that weren't designed for it.
Timestral is Mistral's answer: small, fast, state-of-the-art, and CPU-friendly. The 209M model achieves what much larger systems can't, and the 28M version is a compelling choice for anyone who needs production-grade forecasting without the cloud GPU bill that usually comes with it.
If the benchmark claims hold under independent review, this lands as one of the more practical open-weight releases of the year — a model you can actually deploy on modest hardware rather than admire from a distance.
The weights are available on Hugging Face now, and the API is live through Mistral's platform.