On October 15, 2026, Mistral AI announced Timestral, which it calls its first family of time-series foundation models. They produce probabilistic forecasts, such as predicting sales or energy demand with a range of likely outcomes, and can be used zero-shot, meaning without training on your own data first.
What to know
- Sizes: 5M, 28M, 123M and 209M parameters, with a context length of 32,768 time steps.
- Output: forecasts across 21 quantiles, with support for past-only and known-future covariates.
- Performance: Mistral says the 209M model ranks first on GIFT-Eval, TIME and fev-bench. This is the company’s own claim, and a technical report is still to come.
- Speed: Mistral reports 10 ms (5M model) and 140 ms (209M model) on a server CPU for a one-day-ahead forecast.
- Availability: on GitHub and Hugging Face, and through Mistral Forge. The license is Modified MIT, according to Mistral.
Why it matters
Forecasting is a different job from chatting with a model, and open, small models can run on ordinary hardware. Most readers will not need this directly, but it shows Mistral widening beyond text. See Mistral, Mistral Large 4 and the model tracker.
Source: Mistral AI. We summarise in our own words; check the source for the full details.