Timestral
Mistral released Timestral, its first family of time-series foundation models in 5M, 28M, 123M and 209M sizes, for zero-shot probabilistic forecasting on CPU or GPU.
Timestral is Mistral's first time-series model family. It uses a patch-based transformer with time attention and variate attention. It takes univariate or multivariate targets plus past and known-future covariates, and outputs 21 quantiles with context up to 32,768 steps. Mistral says the 209M model leads zero-shot results on GIFT-Eval, TIME and fev-bench, and each size beats other models of similar size.
- Date
- Thursday 15 October 2026
- Lab
- Mistral AI
- Kind
- open-weights
- Access
- open weights (restricted license)
Figures
| Measure | Value | Measured by |
|---|---|---|
| GIFT-Eval MASE (209M) | 0.658 lower is better; 123M 0.659, 28M 0.678, 5M 0.705 | company |
| GIFT-Eval CRPS (209M) | 0.446 lower is better | company |
| TIME MASE (209M) | 0.626 lower is better | company |
| fev-bench SQL skill score (209M) | 50.39 higher is better | company |
| fev-bench MASE skill score (209M) | 39.49 higher is better | company |
Released under a Modified MIT license. Mistral says fine-tuning is available through Mistral Forge.
Sources
- mistral.ai/news/timestral/
- huggingface.co/mistralai/Timestral-123M
- huggingface.co/mistralai/Timestral-28M
This record was checked against its sources on 15 October 2026. How we check