Mistral ships Timestral, time-series models that skip the GPU
Four open-weight models forecast in probabilities rather than single numbers, and the largest, at 209 million parameters, still runs on a CPU.
Mistral released Timestral on 15 October, its first family of time-series foundation models. There are four sizes: 5M, 28M, 123M and 209M parameters, licensed under a Modified MIT licence, and all are built to run on CPU rather than GPU.
Timestral takes a context of up to 32,768 time steps and returns probabilistic forecasts across 21 quantiles, not a single predicted value, with support for multivariate targets and both past-only and known-future covariates. On a standard server CPU using 8 threads, forecasting a day ahead from three weeks of 15-minute history takes 10ms with the 5M model and 140ms with the 209M model. Mistral reports Timestral ranks first on the GIFT-Eval, TIME and fev-bench forecasting leaderboards. The weights are on GitHub, fine-tuning runs through Mistral Forge, and the models can be deployed on-prem, in a VPC, or air-gapped.
Why it matters: forecasting demand, capacity or anomalies has usually meant reaching for classical statistical models or a metered vendor API. An open, CPU-only model that answers in double-digit milliseconds and can be fine-tuned on your own data changes that calculus, especially for anyone who can’t ship customer data to a third party.
The caveat: the benchmark rankings are Mistral’s own, and the full technical report hasn’t been published yet.