Mistral releases Timestral on October 15: 209M probabilistic forecasting model beating larger models on CPU
Mistral released Timestral on October 15: a probabilistic forecasting model with only 209M parameters that outperforms larger models on multiple CPU forecasting benchmarks; aimed at time-series forecasting and decision support with low latency and on-prem deployment.
Mistral released Timestral on October 15, a probabilistic forecasting model with only 209M parameters designed for time-series forecasting and decision-support scenarios, with low latency and on-prem deployment as core design goals. Mistral said Timestral beats larger probabilistic forecasting models on multiple CPU benchmarks covering standard datasets in energy, transportation, retail and finance. Timestral produces probabilistic outputs rather than point predictions, giving distributional estimates of outcomes for risk-sensitive decisions in uncertain environments; the model weights are open-source and run directly on CPU and edge devices without GPU acceleration, with single-machine latency significantly lower than GPU-based large probabilistic models. Mistral published detailed model cards, probability calibration evaluations, comparisons against GARCH, ETS and Transformer baselines and Python inference and fine-tuning examples; target customers include energy dispatch, demand forecasting, inventory management and financial derivatives pricing.