[New Model] TiRex-2 zero-shot multivariate forecasting

Is your feature request related to a current problem? Please describe. @dennisbader, good day! It’s been a while since we last spoke. Here’s my first issue here as an NXAI engineer:

NXAI recently released TiRex-2, a xLSTM-based foundation model designed for multivariate forecasting. Compared to TiRex-1, it offers several advantages:

• It natively supports multivariate forecasting with covariates.
• It outperforms TiRex-1 across GIFT and fev benchmarks, securing top-3 rankings.
• It maintains high efficiency in terms of memory footprint and speed.
• It relaxes its license to Apache 2.0, making it compatible with commercial use.

We at NXAI believe that TiRex-2 could greatly benefit Darts users and would like to contribute its support to the Darts foundation model stack. We have already implemented a local integration that works seamlessly.

Describe proposed solution

Add a TiRex2Model wrapper for TiRex-2, integrated with Darts' foundation model API, providing:

• Zero-shot forecasting capabilities for both univariate and joint multivariate targets.
• Support for past and future covariates through the standard Darts interface.
• Default median forecasts, with the option to generate quantile forecasts and probabilistic sampling through QuantileRegression.
• The ability to load pretrained checkpoints from Hugging Face or a local directory.
• Seamless integration with existing prediction and backtesting workflows.

We recommend keeping TiRex-2 as an optional dependency and retaining the existing TiRexModel. The initial integration can focus on inference, leaving fine-tuning as an optional setup.

Describe potential alternatives A clear and concise description of any alternative solutions or existing features that might solve it.

Additional context @dennisbader raised a dependency issue on the TiRex-2 repository: https://github.com/NX-AI/tirex-2/issues/1. Since then, we have relaxed the dependencies, ensuring compatibility with Darts’ dependencies.

This could be another sub issue of #2933

• TiRex-2 Paper: https://arxiv.org/abs/2607.01204
• TiRex-2 Repo: https://github.com/NX-AI/tirex-2
• tirex-2 package: https://pypi.org/project/tirex-2/
• TiRex-2 Model Weights: https://huggingface.co/NX-AI/TiRex-2