Mistral's Timestral Beats Larger Rivals at Zero-Shot Forecasting on CPUs

Mistral AI has released Timestral model card, a family of four time-series foundation models with 5 million to 209 million parameters. Mistral reports that the flagship checkpoint leads pretrained zero-shot models on GIFT-Eval, TIME, and fev-bench while supporting CPU inference, giving developers a local forecasting option that requires no model training for each series.

Four sizes, one forecasting interface

Timestral processes numerical sequences in fixed-size patches. The 209M checkpoint groups 32 time steps into each patch and applies attention across both time and variates, meaning parallel signals or columns, through 24 transformer layers with a 768-dimensional embedding.

The largest checkpoint accepts up to 32,768 historical steps and predicts as many as 1,024 steps in one forward pass. For longer horizons, it generates forecasts in blocks and feeds each block’s median prediction back into the context. That recursive approach extends the horizon, although errors can accumulate between blocks.

Each forecast contains 21 quantiles from 0.01 to 0.99, allowing applications to calculate prediction intervals and other uncertainty measures. Timestral accepts univariate or multivariate inputs, past-only covariates, and known-future inputs such as calendar flags or planned prices. It automatically target-encodes categorical variables and represents missing observations as NaN.

Smaller checkpoints challenge larger rivals

On the results published by Mistral, Timestral-209M recorded a mean absolute scaled error of 0.658 and a continuous ranked probability score of 0.446 on GIFT-Eval. MASE measures point-forecast accuracy relative to a baseline, while CRPS evaluates the full predicted distribution. Lower values are better for both metrics.