Mistral's Timestral forecasts numbers, and runs on a plain CPU
On 15 October, Mistral released Timestral, its first family of time-series foundation models. A time series is any number recorded over time, such as hourly sensor readings, daily sales or calls per hour. A foundation model for it is trained on many different series, so it can forecast data it has never seen without being trained on your data first. Mistral calls this forecasting zero-shot.
The models are small by today's standards. There are four sizes, from 5 million to 209 million parameters, and Mistral says none of them needs a GPU to make forecasts.
What Mistral says it does
Mistral says the 209 million parameter model ranks first on the three main public forecasting benchmarks, GIFT-Eval, TIME and fev-bench, and beats the previous best with about two thirds of the parameters. It also says each of the four sizes forecasts better than any public model of its size, and that the 28 million parameter model beats Chronos-2, a model about four times larger, on two of those benchmarks. Mistral has submitted the results to the GIFT-Eval and TIME leaderboards.
Forecasts come as ranges, with 21 quantiles, so you can see how uncertain a prediction is instead of getting a single number. The models can forecast several related series together and take in extra inputs that are known ahead, such as a holiday calendar, a production schedule or a weather forecast.
How fast it is
On a standard server CPU with 8 threads, Mistral says forecasting one day ahead from three weeks of 15-minute data takes 17 milliseconds with the smallest model and 113 milliseconds with the largest. The 123 million parameter model scores within 0.2% of the largest while taking three quarters of the time.
Mistral's own example is an industrial plant with thousands of sensors, where Timestral flags unusual readings and a language model agent then digs through maintenance records to suggest a cause.
Where to get it
Mistral published a GitHub repository for running Timestral on your own data, under what it lists as a modified MIT license. Companies that want to fine-tune it on their own data can do that through Mistral Forge, which can run on their own hardware, including machines with no internet connection. A technical report with full results is still to come.
What to try
If you already forecast something in a spreadsheet, such as weekly orders, staffing for busy hours or stock levels, export a year of that history and run it through the smallest Timestral model from the repository. Compare its forecast for the last month with what actually happened, and with the method you use today. Read the license terms before you put it into a product.
What we think
Most of the AI news this year is about very large models. A forecasting model that runs in milliseconds on an ordinary server is a different kind of useful, because forecasting is work small teams already do by hand every week. We would wait for the technical report before trusting the benchmark ranking, and test on our own numbers either way.
Sources