Mistral AI launches Timestral time series forecasting models

French AI company Mistral AI released Timestral on October 19, 2026, a family of foundation models built for time series forecasting. The lineup spans several sizes, from a compact 5 million parameter version up to a 209 million parameter flagship.

According to the release details, the 209M Timestral model ranks first among zero-shot models on three benchmarks: GIFT-Eval, TIME and fev-bench. Zero-shot here means the model produces forecasts on new data without being fine-tuned for that particular dataset.

Taking on Google's TimesFM-3

Timestral is reported to outperform TimesFM-3, Google's time series foundation model, on these evaluations. The TimesFM line is among the best-known models in the field. The available information does not include a detailed score breakdown, nor does it indicate whether the results have been independently reproduced.

Another notable trait is that the model runs on a standard CPU rather than requiring a GPU. That could lower the cost of deployment for organisations and developers without access to specialised accelerator hardware.

Why forecasting models matter

Time series forecasting underpins predictions about anything measured over time, such as electricity demand, retail sales, inventory needs, financial indicators and weather. Foundation models in this area aim to replace the practice of training a separate model for every dataset with a single general-purpose system.

Mistral AI is best known for its large language models, so Timestral marks a move into specialised models for non-text data. Its small footprint and CPU support may make it easier to embed in business systems, though how it performs on real-world workloads will become clearer as users put it to the test.