Chronos-2
Release in the Chronos (time series models) family · version Chronos-2 (amazon/chronos-2, 120M)
Maintained by Amazon (Amazon Web Services)187
Chronos-2 is a 120M-parameter, encoder-only time-series foundation model from Amazon, released on October 20, 2025. Inspired by the T5 encoder, it produces multi-step quantile forecasts and uses group attention to learn in context across related series and covariates, supporting univariate, multivariate, and covariate-informed forecasting with contexts up to 8,192 steps and horizons up to 1,024 steps.13
- Model hub: Model card (Hugging Face) (external site: huggingface.co)
- Repository: chronos-forecasting repository (external site: github.com)
- Paper: Chronos-2: From Univariate to Universal Forecasting (arXiv 2510.15821) (external site: arxiv.org)
- Release notes: Amazon Science blog post (external site: amazon.science)
Availability and license
Model-disclosure tier
The model parameters for this release can be downloaded by the public. License terms may still restrict use, redistribution, or commercial use.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Public | Weights (safetensors) are published on Hugging Face without an access gate.2 |
| Inference codeIs code for running the model published? | Public | The Apache-2.0 chronos-forecasting package (version 2.0 or later) provides the Chronos2Pipeline used in the card's examples.136 |
| Training codeIs the code used to train the model published? | Unknown | The repository's training scripts are documented for the original Chronos models of March 2024; training code for Chronos-2 was not found.5 |
| Training-data informationPublic = the training data itself can be obtained. Partial = composition or sources are documented without full access. | Partial | The card lists subsets of the public Chronos Datasets and GIFT-Eval Pretrain collections plus synthetic univariate and multivariate data; the report describes how the synthetic data was generated, but the synthetic data itself is not documented as released.17 |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Partial | The technical report describes the architecture, training, and training data at a summary level.7 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Partial | The report and card give results on fev-bench, GIFT-Eval, and Chronos Benchmark II; the repository announces the separately released fev benchmarking package, but instructions to re-run all reported Chronos-2 evaluations were not verified.713 |
What it is useful for
Run and use notes
- The card documents local installation with pip ("chronos-forecasting>=2.0") and a pandas-based prediction API, and deployment to SageMaker with AutoGluon-Cloud or SageMaker JumpStart.1
Organization context
Provenance and derivatives
Other releases in the Chronos (time series models) family
No other releases in this family have been assessed.
U.S. eligibility
Eligible · basis: Documented U.S. control
The model is published by Amazon on Hugging Face and GitHub, and the Chronos-2 report and Amazon's science blog attribute it to Amazon Web Services, a segment of Amazon.com, Inc. Amazon.com, Inc. has its principal executive offices in Seattle, Washington, per its fiscal 2025 Form 10-K.1789
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.