This repository provides the pre-trained forecasting model TiRex introduced in the paper TiRex: Zero-Shot Forecasting across Long and Short Horizons with Enhanced In-Context Learning.

TiRex Model

TiRex is a 35M parameter pre-trained time series forecasting model based on xLSTM.

Key Facts:

• Zero-Shot Forecasting: TiRex is a pre-trained model that performs time series forecasting without requiring any training on your data. Simply download and use it.

• Quantile Predictions: TiRex provides both point estimates and quantile estimates.

• State-of-the-art Performance over Long and Short Horizons: TiRex achieves top scores in various time series forecasting benchmarks, see GiftEval and ChronosZS. These benchmark show that TiRex provides great performance for both long and short-term forecasting.

Documentation

A detailed documentation of TiRex can be found here: https://nx-ai.github.io/tirex/

Installation

pip install tirex-ts

Install with additional input/output adapter:

pip install "tirex-ts[gluonts,hfdataset]"

You can also install TiRex with all extras at once using:

pip install "tirex-ts[all]"

TiRex is currently only tested on Linux and MacOS.

Quick Start

import torch
from tirex import load_model, ForecastModel

model: ForecastModel = load_model("NX-AI/TiRex")
data = torch.rand((5, 128))  # Sample Data (5 time series with length 128)
quantiles, mean = model.forecast(context=data, prediction_length=64)

We provide an extended quick start example in examples/quick_start_tirex.ipynb. This notebook also shows how to use the different input and output types of you time series data. You can also run it in Google Colab.

We provide notebooks to run the benchmarks: GiftEval and Chronos-ZS.

TiRex Classification Model

For detailed instructions on using TiRex classification model please visit our documentation page and quick start Notebook.

TiRex Regression Model

For detailed instructions on using TiRex regression model please visit our documentation page or our quick start Notebook.

TiRex Docker image

For detailed instructions on building and running TiRex in a Docker container, see the Docker README or our deployment documentation.

Finetuning TiRex

TiRex already provide state-of-the-art performance for zero-shot prediction. Hence, you can use it without training on your own data.

If you are interested in models fine-tuned on your data or with different pretraining, please contact us at [email protected]

CUDA Kernels

Tirex can use custom CUDA kernels for the sLSTM cells. These CUDA kernels are compiled when the model is loaded the first time. The CUDA kernels require GPU hardware that support CUDA compute capability 8.0 or later. We also highly suggest to use the provided conda environment spec. The CUDA kernels are automatically used when the xlstm package is installed.

To install TiRex with the CUDA kernels run:

pip install "tirex-ts[cuda,gluonts,hfdataset]"

Explicitly set the custom CUDA backend:

model = load_model("NX-AI/TiRex", backend="cuda")

Troubleshooting CUDA

This information is taken from the xLSTM repository - See this for further details:

For the CUDA version of sLSTM, you need to specify Compute Capability, see https://developer.nvidia.com/cuda-gpus. Or just specify a range as in the example below:

export TORCH_CUDA_ARCH_LIST="8.0;8.6;9.0"

For all kinds of custom setups with torch and CUDA, keep in mind that versions have to match. Also, to make sure the correct CUDA libraries are included you can use the XLSTM_EXTRA_INCLUDE_PATHS environment variable now to inject different include paths, for example:

export XLSTM_EXTRA_INCLUDE_PATHS='/usr/local/include/cuda/:/usr/include/cuda/'

or within python:

import os
os.environ['XLSTM_EXTRA_INCLUDE_PATHS']='/usr/local/include/cuda/:/usr/include/cuda/'

Cite

If you use TiRex in your research, please cite our work:

@inproceedings{auer:25tirex,
  title = {{{TiRex}}: {{Zero-Shot Forecasting Across Long}} and {{Short Horizons}} with {{Enhanced In-Context Learning}}},
  author = {Andreas Auer and Patrick Podest and Daniel Klotz and Sebastian B{\"o}ck and G{\"u}nter Klambauer and Sepp Hochreiter},
  booktitle = {The Thirty-Ninth Annual Conference on Neural Information Processing Systems},
  year = {2025}
  url = {https://arxiv.org/abs/2505.23719},
}

@inproceedings{auer:25tirexclassification,
    title = {Pre-trained Forecasting Models: Strong Zero-Shot Feature Extractors for Time Series Classification},
    author = {Andreas Auer and Daniel Klotz and Sebastinan B{\"o}ck and Sepp Hochreiter},
    booktitle = {NeurIPS 2025 Workshop on Recent Advances in Time Series Foundation Models (BERT2S)},
    year = {2025},
    url = {https://arxiv.org/abs/2510.26777},
}

License

TiRex is licensed under the NXAI community license.