Mistral AI has moved into forecasting. The French lab has released Timestral, its first family of time-series foundation models, with sizes ranging from 5 million to 209 million parameters.
Time-series models predict numbers over time: sales figures, sensor readings, electricity demand, website traffic. Timestral forecasts zero-shot, meaning it can make predictions on data it has never seen without being retrained for each new task.
Mistral reports that the 209M model leads all pretrained zero-shot models on three benchmark suites — GIFT-Eval, TIME and fev-bench — ahead of Google’s TimesFM-3. On fev-bench it scored 50.39 against TimesFM-3’s 48.99. Independent testing has not yet confirmed those results.
The practical hook is efficiency. Timestral runs on a standard server CPU with no GPU required, and ships with open weights and inference code under a modified MIT license. The models output 21 quantiles per time step, producing probabilistic forecasts rather than single-point guesses, and handle context windows of up to 32,768 steps.
That combination — tiny by frontier-model standards, open, and CPU-friendly — makes forecasting one of the cheapest AI workloads to run in production. Analysts, engineers and operations teams can experiment locally instead of provisioning expensive accelerators.
The release follows the same open-weight playbook that popularized models like DeepSeek, and it is Mistral’s latest expansion beyond chat: the company’s assistant Le Chat remains its flagship consumer product.
Source: MarkTechPost
