Mistral's Timestral forecasting AI runs on a server CPU
Mistral's forecasting AI runs without a GPU
Mistral AI has released a forecasting model family that runs on a standard server CPU, no GPU needed.
Timestral is Mistral's first family of time series foundation models. It has four zero-shot forecasting models, from 5M to 209M parameters. The weights and inference code are open. By Mistral's own reported numbers, the largest model ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench. On GIFT-Eval its lead over t0-gamma is only 0.003.
Speed is the practical part. On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history took 10 ms for the 5M model and 140 ms for the 209M model. Scaling flattens at the top: the 123M model scores 0.659 MASE on GIFT-Eval against 0.658 for the 209M, at three quarters of the latency.
Timestral is a causal, patch-based transformer. Context runs up to 32,768 steps, and each model outputs 21 quantiles per step. The weights use a modified MIT licence, while TimesFM-3 is non-commercial. Mistral lists the 209M model as Public Preview.
The scores come from Mistral's model card, and Datadog's card gives Toto-2.0 different TIME scores than Mistral's does.
Would you trust a forecast from a model that runs on a plain server CPU?