Mistral releases Timestral for forecasting sensor data
The model family uses related measurements and known future inputs, with CPU inference available.
Mistral announced Timestral on October 15, 2026, for time series: measurements recorded over time. Its industrial example combines a pump's vibration, temperature and pressure instead of treating each signal separately. The company describes predictions that express uncertainty through quantiles, points in the predicted distribution, rather than one expected value. Production schedules can supply known future inputs. Mistral cautions that operational data can differ from benchmark data and change over time.
The official repository provides standalone inference for CPU or GPU, specifying Python 3.10 or later and PyTorch 2.4 or later. Developers load a model with from_pretrained, which downloads its configuration and weights once into the Hugging Face cache. They then supply numerical history and choose how many future steps to predict. History can be a NumPy array or PyTorch tensor; NaN marks missing values. Only the latest 32,768 steps are retained when history exceeds that limit. Returned results include both the median forecast and quantiles for examining the predicted distribution.
Mistral's model catalog currently labels Timestral 1.0 209M as Public Preview and links to its downloadable weights.
MarkTechPost also covered the release in its report.