Mistral releases small forecasting models that can run on a CPU
Timestral brings downloadable, probabilistic time-series forecasting to CPU-sized models, with separate terms for the code and weights.
Mistral’s October 15 Timestral release comprises four models, from five million to 209 million parameters.
They forecast numerical sequences rather than generate prose and can run through PyTorch on a CPU or GPU. [1][2]
The models accept multiple related series and additional variables, including information known about the future, such as a planned schedule. They return 21 forecast quantiles rather than only one predicted line. That is useful for testing decisions against a range of outcomes, provided the forecasts are calibrated on the reader’s own data. [2][3]
The code is Apache 2.0. The weights’ Modified MIT licence excludes use under that licence if the company or employer’s global consolidated revenue exceeded $20 million in the preceding month; a separate commercial licence is an option. Before adoption, check those terms and compare forecasts with a simple baseline on held-out periods. Launch benchmarks remain vendor-reported. [1][2][4]
Take it further
Inspect the model and weight terms The model card links installation and weight terms. The weight licence has a $20 million preceding-month company-revenue threshold; review eligibility before use.
Sources & details 4 sources
Discuss in your voice app
Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.
Paste the brief into your voice app to discuss this reporting.Reporting sources
- Mistral · Timestral launch
October 15 release announcement and vendor benchmark claims.
- Mistral · Source code and examples
Original repository documents CPU/GPU inference, covariates, sizes and Apache 2.0 code licensing.
- Mistral · 209M model card
Original model card specifies multivariate inputs, 21 quantiles and numerical forecast outputs.
- Mistral · Weight licence
Full licence opened October 16: Modified MIT terms include a restriction tied to consolidated company revenue. Code and weight licences are distinct.
Images & figures
AI-generated editorial illustration of observed history giving rise to uncertain forecasts, with known future context; not documentary evidence or a measured diagram.
Euangelion · AI-generated editorial illustration · Source · Usage terms