Mistral AI has released Timestral, a new family of four time series foundation models ranging from 5M to 209M parameters. The largest 209M model ranks first among pretrained, zero-shot models on major benchmarks including GIFT-Eval, TIME, and fev-bench, outperforming Google's TimesFM-3. Notably, the model achieves this high performance without requiring a GPU at inference.
The Timestral models are designed for developers seeking efficient probabilistic forecasts on standard server CPUs, with open weights and inference code available on GitHub. The architecture uses a causal, patch-based transformer with time attention and variate attention, supporting context windows up to 32,768 time steps. Performance tests show the 209M model completing a 1-day forecast from 3 weeks of 15-minute history in 140 milliseconds using 8 threads on a standard server CPU.
This release matters because it lowers the barrier for developers to run advanced zero-shot time series forecasting efficiently on commodity hardware. By providing open-source models that rival or beat larger competitors while operating solely on CPUs, Mistral AI expands accessible forecasting capabilities for enterprise applications.