Mistral AI Releases Timestral, a Model That Forecasts Future Numbers — and You Can Download and Run It Right Away

On October 15, Mistral AI unveiled Timestral, its first family of time series forecasting models. Time series forecasting is the practice of reading numbers recorded in chronological order and predicting what comes next. Timestral produces future values for multiple variables at once. The family comes in four sizes, and the trained parameter files (model weights) and the code to run them have been available for download since launch day.

21 predictions instead of one, and why that matters

Timestral does not hand back a single number for the future. For every forecast point, it also returns 21 quantiles that describe the distribution of possible future values. A quantile is a reference point that tells you the probability that the actual value will fall below it. In other words, alongside a headline forecast, you also get a read on how much uncertainty surrounds that future value. The model can take in not only past records but also information known in advance, such as a future schedule.

The core design goal is to cut down on the burden of retraining a model from scratch every time the data changes. Long established statistical methods such as ARIMA are typically fitted separately to each individual series. Timestral is built so that a pre-trained model can be applied to new data as is. Mistral AI said it trained the models largely on computer-generated synthetic data, and that it kept the component that reads past records separate from the component that generates future values.

First place in benchmarks, but compared against whom

In its announcement, Mistral AI said that Timestral-209M, the largest model in the family, took first place on all five comparison metrics across three benchmarks. The comparison set consists of models that were pre-trained and then evaluated without any additional fine-tuning on the evaluation data. The three benchmarks are GIFT-Eval, TIME, and fev-bench. On the same scorecard, the company said its model outperformed TimesFM-3 and Chronos-2 on forecasting accuracy.

Speed varied with model size. Reading three weeks of records logged at 15-minute intervals and forecasting one day ahead finished in the blink of an eye. In the company’s own measurements, using 8 concurrent processing units (threads) on a server-class CPU, the smallest model took 0.01 seconds and the largest took 0.14 seconds. The released models run on both CPU and GPU, so you can pick based on the hardware you have.

There are limits on how much the models can handle at once. They reference up to 32,768 past data points, and a single pass produces future values for up to 1,024 points. Longer horizons are forecast in chained segments.

Can companies in Korea use it right away?

Users in Korea can download the models and code from the public repository and run them locally. Enterprises can inquire through Mistral Forge about fine-tuning on their own data and on-premises deployment. This is not a chatbot style question-and-answer product; it is squarely focused on forecasting numerical observations recorded over time.

The usage terms differ between the code and the models. The code carries an Apache 2.0 license, while the model weights carry a modified MIT license. If your own company or your employer had worldwide consolidated revenue exceeding $20 million in the preceding month, you cannot use the weights under that license, and must instead request a separate commercial license or use Mistral AI’s hosted service. Timestral widens the range of forecasting models that can be applied directly to new data. Anyone considering adoption should weigh accuracy on their own data, speed on their specific hardware, and the licensing terms together.