Mistral released Timestral on 15 October 2026: four open models that forecast sales, energy use or sensor readings without first training on your data. In Mistral’s own benchmark table, the largest version beats Google’s TimesFM-3 and Amazon’s Chronos-2, and it runs on a standard server processor. The catch is in the licence: a company with more than $20 million in monthly revenue may not use the models without a separate deal with Mistral.
Timestral comes in four sizes, from 5 million to 209 million parameters. That is small enough to run without a graphics card (GPU). You give it up to 32,768 past time steps, for example three weeks of 15-minute meter readings. It returns a range of likely outcomes instead of one number: 21 estimates, from a 1 percent to a 99 percent level. It can handle several related series at once, plus extra inputs whose future values you already know, such as a holiday calendar.
Mistral says the smallest model makes a one-day forecast in 10 milliseconds on an eight-thread server CPU, and the largest in 140 milliseconds.
Timestral-209M beats TimesFM-3 and Chronos-2 by a few percent on Mistral’s table
On the three public benchmarks Mistral reports, its largest model is ahead of the best-known alternatives, but not by much.
| What is measured | Timestral-209M | TimesFM-3 (Google) | Chronos-2 (Amazon) |
|---|---|---|---|
| Forecast error on GIFT-Eval (MASE, lower is better) | 0.658 | 0.667 | 0.698 |
| Fit of the forecast range on GIFT-Eval (CRPS, lower is better) | 0.446 | 0.456 | 0.485 |
| Forecast error on TIME (MASE, lower is better) | 0.626 | 0.640 | 0.662 |
| Overall skill on fev-bench (SQL skill score, higher is better) | 50.39 | 48.99 | 47.28 |
Mistral ran these tests and lists the rival scores on its own model card. MarkTechPost noted that Google and Amazon do not publish matching figures, and Mistral’s technical report is still to come. The smaller models stay close: the 28M version beats Chronos-2 on GIFT-Eval, and the 123M version scores within 0.2 percent of the largest while taking about a quarter less time.
The modified MIT licence shuts out companies above $20 million a month
Mistral calls the licence “modified MIT”, and the modification is a big one. You may not use the models if your company, or your employer, had global consolidated revenue above $20 million in the previous month. That is roughly $240 million a year. The rule also covers fine-tuned versions and products built on Timestral, and it applies whether you get the model from Mistral or from someone else.
Companies above the line can ask Mistral for a commercial licence, which Mistral may grant at its own discretion. It publishes no price. That hits the sectors Mistral itself names as use cases: energy, supply chains, maintenance and finance are full of companies above that revenue line.
The alternatives have different terms. Amazon’s Chronos-2 is Apache 2.0 licensed, which allows commercial use at any company size, and its error on GIFT-Eval is about 6 percent higher than Timestral-209M’s in Mistral’s table. Google’s TimesFM-3 carries a non-commercial licence, MarkTechPost reported.
Timestral turns a first forecast into a same-day test, if the licence allows it
Forecasting demand, stock levels or energy use usually means a data scientist building and maintaining a separate model for each dataset. For the organisations DataNorth works with, a model that forecasts without training changes the first step. You can run a forecast on your own history the same day, inside your own network, without buying GPUs. That makes it a realistic starting point for planners and controllers, not only for data science teams. It does not replace knowing your data: promotions, maintenance stops and holidays still have to go in as inputs before the forecast can account for them.
Check the licence first. If your company, including the group it belongs to, stays below the threshold, download the 28M or 123M model from the Timestral collection on Hugging Face and backtest it: let it forecast a period you already know and compare the result with your current method. Mistral lists Timestral as a public preview, so expect changes. Above the threshold, talk to Mistral before you build anything, and test Chronos-2 alongside it so you keep a free option if the terms do not work out.
For more information, visit the official announcement of Timestral on the Mistral blog.
