Mistral Releases Timestral, a CPU-First Time Series Forecasting Family
Mistral has introduced Timestral, which it calls the first family of models built specifically for probabilistic multivariate time series analysis. The line has four configurations from 5M to 209M parameters, and the largest one tops the main time series forecasting benchmarks, ahead of the bigger Chronos-2 and TimesFM-3.
The architecture is tuned for CPUs rather than GPUs: the 209M model produces a one-day forecast in 140 ms on eight threads, and the 5M model does it in 10 ms. Context covers up to 32,768 timesteps, and the output is a probability distribution across 21 quantiles rather than a single point estimate.
Weights and code are published under a modified MIT license, with enterprise deployment and fine-tuning on private data offered through Mistral Forge.
Related: Google Open-Sources TimesFM-3 Time Series Forecasting Model