Timestral: AI that doesn't write. It forecasts.

On October 15, 2026, Mistral AI introduced Timestral, a family of models for time series: measurements collected over time, such as energy consumption, temperature, orders or production volumes.

Timestral is not a chatbot and is not designed to generate prose. Its job is to examine how one or more quantities have changed and estimate what may happen next.

Its main feature is *zero-shot forecasting*: a pretrained model can generate an initial forecast for a new dataset without being trained from scratch on that company's historical records.

The family includes 5M, 28M, 123M and 209M-parameter models. It supports up to 32,768 time steps of context and produces 21 quantiles, expressing a range of outcomes rather than just one predicted number.

Timestral can model several related variables together and use information known in advance, such as production schedules or planned shutdowns. The official repository documents both historical inputs and external covariates.

Mistral says inference does not require a GPU. In its published test, forecasting the next 24 hours from three weeks of 15-minute observations takes 10 to 140 milliseconds on an eight-thread server CPU, depending on model size. These are test-environment measurements, not guarantees for every machine.

In its technical announcement, Mistral reports first-place results among pretrained, zero-shot models on GIFT-Eval, TIME and fev-bench. The comparison is based on particular public datasets and metrics.

Those results are notable for a relatively compact model, but they do not establish that Timestral will always outperform classical statistical methods or forecasting systems already tuned to a particular company.

Before operational use, teams should evaluate it on held-out historical periods, compare it with a simple baseline and measure errors, compute costs and forecast stability.

Consider a factory recording machine-level electricity consumption every fifteen minutes. A model could estimate the next day's load from the historical pattern and, when available, shift plans or production schedules.

The same approach could be tested for demand, inventory, order volumes and sensor measurements such as temperature, vibration or pressure. Probabilistic forecasts can also help highlight readings that fall outside expected ranges.

An unexpected reading, however, does not identify a failed component by itself. Maintenance history, operating conditions, production events and human investigation are still needed. Forecasting a measurement is not the same as determining the correct action.

Mistral released the inference code on GitHub under Apache 2.0 and the model weights on Hugging Face under a separate Modified MIT license. Both sets of terms should be reviewed before use.

The models run on CPUs or GPUs and can load files locally. This makes it possible to test forecasting on private infrastructure without necessarily sending production data to a cloud service.

Operational value depends on reliable input data, real-world constraints, understandable thresholds and people accountable for decisions. Timestral can be one component of that system; it is not the system itself.