Chronos-2

Chronos-2 is Amazon's 120-million-parameter encoder-only time-series foundation model for zero-shot forecasting, inspired by the T5 encoder with a group-attention mechanism for in-context learning across related series. It handles univariate, multivariate, and covariate-informed tasks in one architecture, with 8K context and 1K prediction length, and Amazon reports state-of-the-art zero-shot accuracy among public models on fev-bench, GIFT-Eval, and Chronos Benchmark II. Weights are Apache 2.0, it runs 300+ forecasts per second on a single A10G, and it is already in production at retailers like Decathlon.

The honest picture: at 120M parameters it is far smaller than Google's TimesFM-3, and univariate-only workloads may see little gain over Chronos-Bolt. The headline benchmarks are Amazon-reported, and production stories like Decathlon's involve fine-tuning with LoRA — pure zero-shot results on your data may differ from the marketing.