Timestral: Mistral's time-series foundation models

Mistral's Timestral models are time-series foundation models designed for efficient inference without a GPU. On a standard server CPU with 8 threads, the 5M model forecasts one day ahead from three weeks of 15-minute history in 10 ms, while the 209M model takes 140 ms. Smaller models like the 28M are five times faster than the 209M model, and the 123M model achieves a score within 0.2% of the 209M model at three quarters of its latency. Timestral is released under the Modified MIT license.