PickTimestral-209M: 32,768 time series data points, 1,024 future predictions

Timestral-209M, released by Mistral AI, is a foundation model that predicts time series data directly without training. With a scale of 209M parameters, it handles both single-variable and multi-variable time series. It distinguishes between inputs where only historical data is available and cases where future covariates are known.

1 / 2

It accepts up to 32,768 historical time steps and predicts up to 1,024 steps in a single forward pass. It outputs a probabilistic distribution of 21 quantiles ranging from 0.01 to 0.99, rather than simple point predictions. Missing values are handled by representing them as NaN, and target encoding is applied to categorical covariates.

On three major benchmarks—GIFT-Eval, TIME, and fev-bench—it recorded lower MASE and CRPS scores than existing models such as t0-gamma, TimesFM-3, and Chronos-2. In particular, it achieved a score of 50.39 on the SQL skill score in fev-bench, outperforming competing models. It is built on PyTorch and supports inference on both CPU and GPU.

While existing time series models require retraining on specific domain data, Timestral captures diverse time series patterns in a zero-shot manner. The source code is open, making it easy to reproduce benchmark results. It is distributed under the Modified MIT license, with checkpoints of various sizes provided, ranging from 5M to 209M.