Mistral's Timestral Forecasts Numbers Over Time, and Runs on a Plain CPU

Mistral AI, the French lab, introduced Timestral on October 15. Unlike the chatbots it is better known for, Timestral does not write text. It is a family of models for forecasting: you feed in a history of numbers, such as electricity use, sales or sensor readings, and it predicts what comes next. Researchers call this “time series” data, simply measurements taken over time.

There are four sizes, from 5 million to 209 million parameters (the adjustable numbers inside a model, so these are tiny next to chatbots). Instead of one guess, Timestral gives a probabilistic forecast across 21 quantiles, which is a fancy way of saying a range: “most likely this, but plausibly anywhere between these two values”. It can take in several related series at once, plus extra information such as known future holidays. Mistral says no GPU is needed. On an 8-thread server CPU, a one-day-ahead forecast from three weeks of 15-minute history takes about 10 milliseconds with the smallest model and 140 milliseconds with the largest.

Mistral reports that the 209M model ranks first among pretrained, zero-shot models on three public forecasting benchmarks (GIFT-Eval, TIME and fev-bench). “Zero-shot” means it works on your data without extra training. A fair caveat: these are Mistral’s own results, and the full technical report is still to come. The smaller 5M and 28M models trade some accuracy for speed. The licence is a modified MIT, and the code and models are on GitHub and Hugging Face. Companies can also fine-tune it and run it on-premises or in air-gapped setups.

What’s behind it: most of the AI headlines are about large language models, but a lot of everyday business is numbers over time: how many customers visit on Friday, how much power a building will use, when a machine is likely to need service. Foundation models for time series apply the same idea as chatbots, training one general model on huge amounts of data so it can handle new tasks without starting over. Small size is the point here. A model this light can sit next to the data instead of in a distant data centre.

What this means for you: for most people, nothing changes today, and that is fine. If you work with spreadsheets of recurring numbers, say a small shop’s sales or a home energy log, this hints at forecasting tools that run on ordinary hardware without sending data away. Developers can try it on a laptop. As always, treat a forecast as a range of possibilities, and check it against your own past data before trusting it.

Sources

• Mistral AI: Introducing Timestral