mldi-lab/Kairos_50m

Kairos-50M: Adaptive Time Series Foundation Model

This model is presented in the paper Kairos: Towards Adaptive and Generalizable Time Series Foundation Models.

Model Description

Kairos-50M is a 50-million parameter time series foundation model designed for zero-shot forecasting across diverse domains. It features adaptive tokenization and instance-specific positional encodings to handle heterogeneous time series data with varying information density.

Key Features

• 🔀 Mixture-of-Size Dynamic Patching (MoS-DP): Adaptively selects tokenization granularity based on local information density
• 🔄 Instance-adaptive Rotary Position Embedding (IARoPE): Tailors positional encodings to unique temporal characteristics of each series
• 📊 Zero-shot Forecasting: Strong generalization across domains without fine-tuning
• ⚡ Efficient: Superior performance with fewer parameters

Model Specifications

• Parameters: ~50 million
• Training Data: PreSTS corpus (300+ billion time points)
• Architecture: Transformer-based with adaptive components

Model Family

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

Usage

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)

For detailed usage examples, please refer to the main repository.

Citation

If you use this model, please cite:

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

License

Apache License 2.0