Kairos: Toward Adaptive and Parameter-Efficient Time Series Foundation Models

📅 News

• 14 May 2026: 📅 Updated Kairos paper (v3) released.
• 16 Feb 2026: 📅 Updated Kairos paper (v2) released.
• 06 Oct 2025: ✨ Kairos is now on the GIFT-Eval Leaderboard.
• 30 Sep 2025: 📅 Kairos paper and inference code released.

🌟 Introduction

Kairos is a flexible and parameter-efficient Time Series Foundation Model (TSFM) designed to handle the dynamic and heterogeneous nature of real-world time series data. Unlike existing models that rely on rigid, non-adaptive processing pipelines and massive parameterization, Kairos decouples temporal heterogeneity from model capacity through three key architectural innovations:

• 🔀 Mixture-of-Size Encoder: Adaptively tokenizes time series at multiple granularities based on local information density. It utilizes a Top-K granularity router with null experts to efficiently model diverse temporal patterns.

• 🔄 Heterogeneity-Aware Transformer: Incorporates Dynamic Rotary Position Embedding (DRoPE), a granularity-aware positional encoding that modulates temporal scales using instance-level spectral features. It adapts to the varying physical durations of dynamic patches, enabling robust modeling of diverse temporal dependencies.

• ⏩ Multi-Patch Decoder: Employs learnable forecast tokens to predict multiple future patches in parallel, mitigating cumulative errors in autoregressive generation and offering flexibility for variable-length prediction horizons.

Trained on the large-scale Predictability-Stratified Time Series (PreSTS) corpus comprising over 300 billion time points, Kairos achieves superior zero-shot forecasting performance with significantly fewer parameters compared to existing methods on both GIFT-Eval and Time-Series-Library benchmarks.

⚙️ Method Overview

Overview of the Kairos architecture, highlighting the Mixture-of-Size Encoder, Heterogeneity-Aware Transformer with DRoPE, and the Multi-Patch Decoder.

📊 Evaluation

Kairos achieves superior performance with fewer parameters on two common zero-shot benchmarks.

• GIFT-Eval

• Time-Series-Library (TSLib)

💻 Usage

Prerequisites

Choose one of the following methods to set up the environment:

Option 1: Install from Source (Recommended) Best for users who want to use the library directly. This ensures tsfm is globally accessible in your environment and resolves path issues automatically.

uv pip install git+https://github.com/foundation-model-research/Kairos

Option 2: Local Setup Best for running demos or modifying the code locally.

git clone https://github.com/foundation-model-research/Kairos.git
cd Kairos
uv sync

Note: If you choose Option 2, please ensure the project root is added to your PYTHONPATH or use sys.path.append in your scripts to avoid ModuleNotFoundError.

Model Setup

Our model weights are available on Hugging Face. You can access them at the following links:

• 50M: https://huggingface.co/mldi-lab/Kairos_50m
• 23M: https://huggingface.co/mldi-lab/Kairos_23m
• 10M: https://huggingface.co/mldi-lab/Kairos_10m

Quickstart

The datasets folder contains a specific sequence segment from the ETTh1 dataset, which is a component of the zero-shot test dataset.

You can run our forecasting demo using the quickstart_zero_shot.ipynb notebook.

Alternatively, you can use the following Python code snippet for a quick start to load the Kairos model and generate a forecast:

import torch
from tsfm.model.kairos import AutoModel

# load model
model = AutoModel.from_pretrained(
    "mldi-lab/Kairos_50m", trust_remote_code=True
)

# forecasting configurations
batch_size, context_length, prediction_length = 1, 2048, 96
seqs = torch.randn(batch_size, context_length)

prediction_length = 96
forecast = model(
    past_target=seqs.clone().detach().float(),
    prediction_length=prediction_length,
    generation=True,
    preserve_positivity=True,
    average_with_flipped_input=True
)

# extract the prediction results
forecast = forecast["prediction_outputs"]
print(forecast.shape)

🤝 Acknowledgements

This repository includes code adapted from Chronos: Learning the Language of Time Series. We thank the authors for their excellent work and open-source contributions. We also extend our gratitude to the creators of the following datasets used in our work:

• LOTSA Data
• Chronos Datasets

⚖️ License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

💬 Contact

We welcome any questions, feedback, or potential collaborations. You can reach us at:

• Kun Feng: fengkun2025@shanghaitech.edu.cn
• Shaocheng Lan: lanshch2024@shanghaitech.edu.cn
• Yuchen Fang: yuchen.fyc@antgroup.com

📝 Citation

If you find Kairos models useful for your research, please consider citing the associated paper:

@article{feng2025kairos,
  title={Kairos: Toward Adaptive and Parameter-Efficient Time Series Foundation Models},
  author={Feng, Kun and Lan, Shaocheng and Fang, Yuchen and He, Wenchao and Lu, Sihan and Gu, Shuqi and Ma, Lintao and Lu, Xingyu and Ren, Kan},
  journal={arXiv preprint arXiv:2509.25826},
  year={2025}
}