Chronos

Chronos reframes probabilistic time-series forecasting as language modelling: continuous values are scaled and quantised into a fixed vocabulary of tokens, after which an off-the-shelf transformer (originally based on T5) is trained with cross-entropy loss over next-token prediction. The original Chronos family was pretrained on a large open corpus complemented by Gaussian-process synthetic data and ranged from 20M to 710M parameters.

Chronos-Bolt is the second-generation patch-based variant: instead of one token per timestep, contiguous patches of observations are encoded together, yielding ~250× faster inference at comparable accuracy. The four Bolt sizes (Tiny → Base) trade speed for capacity along a fairly clean Pareto curve.

Chronos-2 (Oct 2025) drops the encoder-decoder structure for an encoder-only design and unifies univariate, multivariate, and covariate-informed forecasting in a single 120M-parameter checkpoint.