Timestral: Mistral's time-series foundation models

Mistral released Timestral, a family of time-series foundation models ranging from 5M to 209M parameters that delivers state-of-the-art zero-shot multivariate forecasting, with the 28M-parameter version outperforming models four times its size on key benchmarks. The 209M-parameter Timestral model beats prior state-of-the-art time-series models on GIFT-Eval, TIME and fev-bench while using fewer parameters, enabling CPU-bound local inference instead of GPU-dependent retraining pipelines for tasks such as capacity planning, equipment monitoring and financial forecasting.

Hacker News https://mistral.ai/news/timestral/ Timestral: Mistral's time-series foundation models Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. Mistral’s Timestral models deliver state-of-the-art zero-shot multivariate forecasting at just 5M to 209M parameters, with the 28M version outperforming models four times its size on key benchmarks. For production engineers, this shifts time-series forecasting from expensive, GPU-dependent retraining pipelines to fast, CPU-bound local inference that can be embedded directly into real-time agentic workflows and telemetry loops. Timestral’s largest model, at 209M parameters, outperforms previous state-of-the-art time-series models on key benchmarks GIFT-Eval, TIME, fev-bench while using fewer parameters, enabling high accuracy with lower computational costs. This allows deployment on CPUs and reduces operational overhead for tasks like capacity planning, equipment monitoring, and financial forecasting, especially in domains with limited data.