What changed
Mistral released Timestral models at 5M, 28M, 123M and 209M parameters for probabilistic multivariate forecasting with past and known-future covariates, 32,768-step context and 21 output quantiles, plus Apache-2.0 inference code and modified-MIT weights.
Why it matters
The release gives forecasting teams a small, inspectable alternative to general-purpose language models and supports uncertainty rather than a single point estimate. Vendor leaderboard claims still need matched backtests because production series, missingness and regime shifts differ from public benchmarks.
What to do
Pin the repository and one model revision, backtest against a seasonal baseline and the incumbent model on rolling windows, measure calibration and coverage for all 21 quantiles, then test CPU latency, missing values and known-future covariates on representative data.
Caveat
Mistral reports first place on GIFT-Eval, TIME and fev-bench among comparable pretrained zero-shot entries, but the result is vendor-run and a full technical report is still forthcoming. The model license differs from the Apache code license, and future covariates must not leak target data.