Mistral has released Timestral, its first family of time-series foundation models, and the pitch is deliberately small. The line runs from 5M to 209M parameters, with the flagship Timestral-209M shipping as an 837MB checkpoint that forecasts on a standard eight-thread CPU in roughly 140 milliseconds. The smallest model does the same job in about 10 milliseconds. No GPU is required for either.
The architectural bet is Decoupled Horizon Forecasting. Instead of masking the values it needs to predict, Timestral uses learned forecast tokens that read the entire history, so every observed value can work as both target and context. Covariate-aware attention then assigns each series a role: target, past-only, or known-future covariate. Information flows only from covariates to targets, which lets a holiday calendar or weather feed help a forecast without leaking the target's own future. Context stretches to 32,768 time steps and each forward pass returns 21 probability quantiles.
Mistral says Timestral-209M ranks first among pretrained zero-shot models on GIFT-Eval at 0.658 MASE, TIME at 0.626, and fev-bench at 50.39 SQL skill. The 28M variant beats Chronos-2, a model roughly four times larger, on the first two. The margins are thin and the results are self-reported until the technical report lands, and the weights ship under a modified MIT license. Even with those caveats, the release points at where small models are heading: narrow, cheap, and able to run anywhere.