Mistral introduces Timestral time-series foundation models
Mistral released Timestral, a family of four time-series foundation models with up to 209M parameters.
Models:TimestralChronos-2TimesFM-3Falcon-2.0Toto 2.0
What changed
Mistral released Timestral, its first family of time-series foundation models. The models come in four sizes: 5M, 28M, 123M and 209M parameters. They support a context length of 32,768 time steps and output probabilistic forecasts across 21 quantiles. Inference runs on standard CPUs without requiring a GPU, with the 5M model taking 10 ms and the 209M model taking 140 ms for a one-day forecast. The models are available via GitHub and Mistral Forge.
Source
Timestral has been optimised as a family of four models: 5M, 28M, 123M and 209M parameters.
Original announcement: Mistral AI news
Source: Mistral AI news (2026-10-15)