Mistral's Timestral forecasts time series and runs on a CPU

Mistral released Timestral, a family of time-series foundation models that forecast future values from past data without being trained on the specific dataset first. There are four sizes, from 5 million up to 209 million parameters, all sharing one architecture. Mistral says the 209M version ranks first among pretrained zero-shot models on GIFT-Eval, TIME, and fev-bench, the three main public benchmarks for this task, and that it does so from a single checkpoint without ensembling. The smaller 28M model reportedly beats Chronos-2, which is about four times larger.

The size matters because these run on ordinary hardware. On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute readings takes about 10 milliseconds with the 5M model and 140 milliseconds with the largest. The technical twist is an approach Mistral calls Decoupled Horizon Forecasting: instead of masking the values it wants to predict, the model keeps the full history in the sequence and uses separate forecast tokens that read everything before them, so each observed value serves as both a target and context. It also tracks which inputs are known-future covariates, like a planned promotion or a scheduled shutdown, and only lets information flow from those into the forecast. Most of the training data is synthetic, and the models are released under a Modified MIT license on GitHub and Hugging Face.

One caveat: the benchmark rankings and speed figures are Mistral's own, and the full technical report is not out yet, so the numbers are not independently verified.

Why it matters

If you forecast demand, sensor readings, or financial flows, a model that runs on a CPU removes the usual GPU cost and latency for this kind of work, and zero-shot use means you can try it on your series before committing to training anything. Wait for the technical report, then test it against whatever statistical baseline or in-house model you run now.