Timestral-123M: Mistral's time-series forecasting model

Mistral has released Timestral, a family of time series foundation models for zero-shot forecasting. Timestral-123M is one of four sizes, alongside 5M, 28M and 209M parameter versions. The models are published on Hugging Face.

The architecture is a patch-based transformer with time attention and variate attention. It takes up to 32,768 time steps of context and forecasts up to 1,024 steps in a single forward pass. Longer horizons are produced block by block, with each block's median forecast appended to the context. Outputs are 21 quantiles, from 0.01 to 0.99.

Inputs can be univariate or multivariate, with past-only and known-future covariates, numeric or categorical. Missing values are marked as NaN. Inference runs in PyTorch on CPU or GPU.

Mistral says Timestral-209M reaches top zero-shot accuracy on GIFT-Eval, TIME and fev-bench, and that every Timestral model outperforms other models of its size. On GIFT-Eval, the 123M version scores 0.659 MASE and 0.448 CRPS, against 0.658 and 0.446 for the 209M model.