We introduce Timestral, a family of four time series foundation models for probabilistic forecasting, with 5M, 28M, 123M, and 209M parameters. Timestral sets a new state of the art on GIFT-Eval, TIME, and fev-bench, where Timestral-209M and Timestral-123M take the top two places on every metric. We…

We introduce Timestral, a family of four time series foundation models for probabilistic forecasting, with 5M, 28M, 123M, and 209M parameters. Timestral sets a new state of the art on GIFT-Eval, TIME, and fev-bench, where Timestral-209M and Timestral-123M take the top two places on every metric. We…

We introduce Timestral, a family of four time series foundation models for probabilistic forecasting, with 5M, 28M, 123M, and 209M parameters. Timestral sets a new state of the art on GIFT-Eval, TIME, and fev-bench, where Timestral-209M and Timestral-123M take the top two places on every metric. We…

We introduce Timestral, a family of four time series foundation models for probabilistic forecasting, with 5M, 28M, 123M, and 209M parameters. Timestral sets a new state of the art on GIFT-Eval, TIME, and fev-bench, where Timestral-209M and Timestral-123M take the top two places on every metric. We…

Introduction: As artificial intelligence transitions from generative text to specialized operational prediction, Mistral AI has introduced Timestral, a lightweight time-series model family optimized for high-efficiency enterprise forecasting. Ranging from 5 million to 209 million parameters, Timestral achieves state-of-the-art performance across temporal benchmarks while maintaining native CPU inference compatibility. In modern cyber defense and infrastructure engineering, leveraging high-speed temporal forecasting allows security teams to anticipate network traffic anomalies, resource exhaustion attacks, and telemetry threats with minimal compute overhead.

Unofficial native Rust inference for Timestral, powered by Candle. CPU and Metal support, quantile forecasts, and an HTTP API - pims/timestral-rs

Mistral launches Timestral, time-series models from 5 to 209 million parameters. They run on CPU and return probability ranges, useful for anticipating load spikes and anomalies in systems.

Mistral AI releases Mistral Timestral, a family of compact time-series foundation models for zero-shot forecasting in four sizes under a modified MIT license.