README.md

metadata

license: apache-2.0
datasets:
  - Salesforce/GiftEvalPretrain
  - autogluon/chronos_datasets
tags:
  - time series
  - time-series
  - forecasting
  - classification
  - anomaly detection
  - imputation
  - foundation models
  - pretrained models
  - time series foundation models
library_name: basicts
paper:
  - https://arxiv.org/abs/2607.01918

Zeus

Zeus is a tuning-free, multi-task time series foundation model for time series analysis. It supports downstream tasks including point forecasting, probabilistic forecasting, imputation, anomaly detection and classification in a tuning-free manner. Trained on a large-scale real-world and synthetic datasets, it achieves state-of-the-art performance among public models on LTSF benchmark, GIFT-Eval, UCR Anomaly Detection Archive, UEA Classification Archive.

🌟Most pretrained models focus solely on forecasting, with other tasks relying on task-specific fine-tuning. In contrast, Zeus is the first to support the following mainstream time series tasks in a tuning-free manner:

Task	Zeus✨	TimesBERT	MOMENT	UniTS	Timer
Long Forecasting	✅	❌	⭕️	⭕️	✅
Short Forecasting	✅	⭕️	✅	⭕️	✅
Probabilistic Forecasting	✅	❌	❌	❌	❌
Imputation	✅	⭕️	⭕️	⭕️	⭕️
Anomaly Detection	✅	⭕️	✅	⭕️	⭕️
Classification	✅	⭕️	⭕️	⭕️	⭕️

✅ Zero-shot ❌ Unsupported ⭕️ Fine-tuned required

Links

📄Paper (Accepted by ICML'26)

📦Github

Quick Start

1. Environment and Dependencies This project is built upon the BasicTS library. Install it with pip:

    pip install basicts>=1.1.0
   

2. Load Model

   Zeus can be loaded from BasicTS library

   from basicts.models.Zeus import ZeusForPrediction # ZeusForImputation, ZeusForClassification
   

   or loaded from this repository

   hf download GestaltCog/zeus
   

   from Zeus.modeling_zeus import ZeusForPrediction # ZeusForImputation, ZeusForClassification
   

   Then load the model weights

   model = ZeusForPrediction.from_pretrained(
       "zeus",
       trust_remote_code=True,
       attn_implementation="flash_attention_2",
       device_map="cuda"
   )
   

3. Run

   prediction = model.generate(torch.rand(1, 256, device="cuda"), prediction_length=16, normalize=True)
   

Citation

If you find Zeus useful, please consider citing this paper:

@inproceedings{fu2026zeus,
title={Zeus: Towards Tuning-Free Foundation Model for Time Series Analysis},
author={Yisong Fu and Zezhi Shao and Chengqing Yu and Yujie Li and Yongjun Xu and Xueqi Cheng and Fei Wang},
booktitle={Forty-third International Conference on Machine Learning},
year={2026}
}

Contact

If you have any questions or want to use the code, feel free to contact: fuyisong24s@ict.ac.cn.