Chronos: versions of the time-series foundation model, and how it works

Chronos is a family of pretrained forecasting models from Amazon. The idea is the same one that made large language models useful: train once on a large collection of series, then forecast a new series in zero shot, without fitting ARIMA or a task-specific network from scratch. The original paper (Ansari and colleagues, 2024) showed that a T5 language-model backbone can do this if the numbers are turned into tokens first. Later releases keep that pretrained-forecasting idea and change how the tokens, patches, and attention are arranged.

The original Chronos models (Hugging Face ids amazon/chronos-t5-tiny through chronos-t5-large) scale a series, usually by its mean, then quantise the values into a fixed vocabulary of 4,096 bins. Those bin ids are what the T5 encoder–decoder sees. The decoder samples future tokens one by one, and the tokens are mapped back to real numbers. That autoregressive loop is why the first generation is relatively slow. The published sizes are tiny (8 million parameters), mini (20 million), small (46 million), base (200 million), and large (710 million). Context is short by later standards: 512 past steps in, 64 steps out.

Chronos-Bolt, released in November 2024, keeps a T5-style backbone but stops forecasting token by token. It chunks the history into patches of several observations, encodes those patches, and lets the decoder emit a multi-step quantile forecast in one go (direct multi-step prediction). Amazon reports about 5 percent lower error than the original models of similar size, up to 250 times faster inference, and about 20 times less memory. The Bolt sizes on Hugging Face are tiny (9 million), mini (21 million), small (48 million), and base (205 million). There is no Bolt-large. Context grows to 2,048 steps; the default forecast length stays 64. Future covariates are not native: AutoGluon can attach an external regressor, which only captures per-timestep effects.

Chronos-2, released in October 2025 as amazon/chronos-2, is a single 120-million-parameter encoder-only model. It still produces quantile forecasts in one forward pass, now over a 21-quantile grid that includes the tails (0.01 and 0.99). The new piece is group attention: related series, variates of a multivariate series, and covariates share information inside a group instead of being concatenated into one long context. That is how Chronos-2 does univariate, multivariate, and covariate-informed forecasting in the same architecture, including past-only and known-future covariates. Maximum context is 8,192 steps and maximum prediction length is 1,024. Amazon reports strong zero-shot numbers on fev-bench, GIFT-Eval, and Chronos Benchmark II; treat those as the authors’ published comparisons, not as a claim I have re-run here.

In practice the working pipeline is: load a pretrained checkpoint (the Python package is chronos-forecasting, and AutoGluon TimeSeries wraps the same models), pass the recent history, and read a predictive distribution rather than a single line. Fine-tuning is available when you have enough in-domain series, but the default is inference-only. Chronos is a forecasting tool. It does not read a medical scan, and a confident quantile band is not a clinical decision. For students, the useful distinction is original Chronos for understanding tokenisation, Bolt when you need speed on a single series, and Chronos-2 when the task is multivariate or has covariates you actually observe.