Timestral 1.0 28M
Timestral 28M is a smaller forecasting checkpoint with 12 layers and a 384-dimensional representation. Mistral positions this size for low-latency forecasts. It supports zero-shot prediction with past-only and known-future inputs.
Preview
Use numeric observations and related variables
The release handles one or multiple target series and uses attention across time and variables. Its numeric output includes 21 quantiles. The published implementation runs in PyTorch on a CPU or GPU. It accepts the same 32,768-step history as the other Timestral sizes.Timestral 28M official model card
Download the forecasting checkpoint
The official model page links to the 28M weights. It does not list a Mistral serverless forecasting identifier. The Modified MIT license needs a separate review from the standard MIT license.Timestral 1.0 28M model page
Compare forecasts on unseen periods
Practical guidance. This model is worth comparing when forecasts are refreshed often or hardware is limited. Evaluate the time to process every required series, not just one short example. Then inspect whether the smaller model loses accuracy on seasonal changes or uncommon peaks that matter to the application.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Preview |
| Access | Downloadable |
| Model identifiers | Timestral-28M |
| Release class | Public preview; downloadable checkpointTimestral 1.0 28M model page |
| Parameters | 28 million nominal parametersTimestral 28M official model card |
| History limit | 32,768 time steps, not language tokensTimestral 28M official model card |
| Single-pass horizon | 1,024 future time stepsTimestral 28M official model card |
| License | Modified MIT; company revenue condition appliesTimestral Modified MIT license |
Related models
- Timestral 1.0 123M
Mistral AI · Timestral
A 123M time-series model with 20 layers that predicts future values and quantiles from numeric observations.
Launch date not confirmedPreviewDownloadable - Timestral 1.0 209M
Mistral AI · Timestral
The 209M Timestral model forecasts numeric series with a 24-layer architecture and quantile outputs.
Launch date not confirmedPreviewDownloadable - Timestral 1.0 5M
Mistral AI · Timestral
The smallest Timestral release uses eight layers and 5M parameters for numeric time-series forecasting.
Launch date not confirmedPreviewDownloadable