Timestral Models by MistralAI
Last refreshed 2026-10-20. Next refresh: weekly.
Details
About
Timestral is Mistral AI's family of time-series foundation models for zero-shot probabilistic forecasting, released October 15, 2026 in 5M, 28M, 123M and 209M sizes. The models handle univariate and multivariate series with past-only and known-future covariates, read up to 32,768 time steps, and output 21 quantiles. Weights are on Hugging Face under Mistral's Modified MIT license; the inference code on GitHub is Apache-2.0.
Decision facts
- Best fit
- forecastingcoding
- Capability starting point
- Timestral 1.0 123M
- Lowest tracked input
- Not tracked
- Closest related family
- Devstral
Current Variants
Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.
Use when the workload needs forecasting and 209M parameters.
Use when the workload needs forecasting and 123M parameters.
Use when the workload needs forecasting and 28M parameters.
Use when the workload needs forecasting and 5M parameters.
| Model | Use when | Released | Signals | Status |
|---|---|---|---|---|
| Timestral 1.0 209M | Use when the workload needs forecasting and 209M parameters. | 2026-10 | forecasting209M parameters | Current |
| Timestral 1.0 123M | Use when the workload needs forecasting and 123M parameters. | 2026-10 | forecasting123M parameters | Current |
| Timestral 1.0 28M | Use when the workload needs forecasting and 28M parameters. | 2026-10 | forecasting28M parameters | Current |
| Timestral 1.0 5M | Use when the workload needs forecasting and 5M parameters. | 2026-10 | forecasting5M parameters | Current |
Release Timeline
1 release groupSpecifications(4 models)
| Model | Released | Parameters |
|---|---|---|
| Timestral 1.0 209M | 2026-10 | 209M |
| Timestral 1.0 123M | 2026-10 | 123M |
| Timestral 1.0 28M | 2026-10 | 28M |
| Timestral 1.0 5M | 2026-10 | 5M |
