The family is available in 5M, 28M, 123M, and 209M parameter versions; the published Hugging Face repository is for the 209M checkpoint.
According to its model card, Timestral supports univariate and multivariate targets, past-only and known-future covariates, and probabilistic forecasts across 21 quantiles. The 209M model accepts up to 32,768 historical time steps and can forecast up to 1,024 steps in one forward pass. Mistral describes it as a 24-layer patch-based transformer with time and variate attention. The repository provides PyTorch usage instructions for CPU or GPU inference.
Mistral reports benchmark results for Timestral-209M on GIFT-Eval, TIME, and fev-bench, including a GIFT-Eval MASE of 0.658 and CRPS of 0.446. Model weights are released under a Modified MIT license; the model card says source code is available in Mistral's GitHub repository. The source does not establish independent reproduction of the benchmark claims, hosted API availability, pricing, or performance on a particular deployment.