Mistral Timestral: what it is and is it free to use?

Timestral is a family of four forecasting models that Mistral released on October 15. You give it a column of numbers over time, such as daily sales, call volume or sensor readings, and it predicts where those numbers are likely to go next. It is free to download and use, commercial use included, as long as your company brings in less than $20 million a month in revenue.

The release spread on X through the paper-tracking account @HuggingPapers, which summed it up as "state-of-the-art zero-shot forecasting across 3 major benchmarks" and pointed people to the weights on Hugging Face. @cheatyyyy noted that this is Mistral's "first family of time-series models." Until now Mistral has shipped chat and coding models. This one doesn't handle language at all, only numbers.

What Timestral actually does

There's no chat box. Timestral is a "foundation model" for time series: Mistral trained it ahead of time on a huge number of series, mostly synthetic ones by its account, so it can forecast data it has never seen without first learning your history. Forecasting people call this "zero-shot."

Here are the specifics from Mistral's announcement and the GitHub repo:

• It comes in four sizes: 5M, 28M, 123M and 209M parameters. The largest file is about 837 MB, tiny next to any chat model.
• It doesn't need a GPU. In Mistral's own test on an 8-thread server CPU, the smallest model forecast a day ahead from three weeks of 15-minute readings in 10 milliseconds. The largest took 140.
• Each forecast comes back as 21 quantiles, so you get a likely value plus a low case and a high case instead of one line.
• You can feed in things you already know about the future, like holidays, promotions or a production schedule, and the model takes them into account.
• It reads up to 32,768 past time steps.

Mistral says the 209M model ranks first among pretrained models on three public forecasting leaderboards (GIFT-Eval, TIME and fev-bench), and that the 28M model beats Amazon's Chronos-2, a model about four times its size. Those are Mistral's own runs. Its docs still label Timestral a public preview, and I haven't seen anyone test it independently yet.

Is Timestral free?

For almost every small business, yes. The inference code is Apache 2.0. The weights use what Mistral calls a Modified MIT license. You can use, modify and even sell work built on them, unless your company's global revenue (or your employer's) topped $20 million in the previous month. Above that line you need a commercial license from Mistral or have to go through Mistral's hosted services. The cap also covers derivatives and fine-tuned versions.

Mistral's docs page doesn't list a hosted API price for Timestral yet. For now you use it through the Python package: install it, load a model from Hugging Face, pass in an array of numbers, get a forecast back. If you're comfortable with Python, I'd guess an afternoon gets you a first forecast.

My read

What caught my attention is the size more than the leaderboard claims. Good forecasting used to mean paying for an analytics product or having a data scientist tune a model for each series. Timestral is a free download that runs on the kind of machine your bookkeeping software already runs on, and it hands you a range you can plan around.

It won't rescue bad data. A shop with six months of messy, half-exported sales figures will still get a forecast that looks confident. The quantile output helps a little here: when the low and high cases sit far apart, the model is telling you it doesn't really know.

What it means for a business

Most owners already forecast in their heads. How many techs to put on next Tuesday, how much stock to order before a holiday, whether cash gets tight in March. A model like this can turn those gut calls into a number with a range around it, refreshed every morning from data you already collect.

Downloading the model takes minutes. Most of the effort goes into pulling clean history out of your POS, scheduling or accounting system, and then getting the forecast to whoever actually makes the call. If you're not sure which of your numbers is worth predicting first, New Face Design's free process audit can help you pick one.