TiRex: NX-AI's xLSTM Zero-Shot Forecasting Model

Most time series foundation models in the current wave are transformer variants. TiRex stands out because it is not. The model introduced by NX-AI is based on xLSTM, positioning it as a recurrent alternative for zero-shot forecasting across both short and long horizons.

The public TiRex release is framed around a simple value proposition: load a pretrained model, provide historical context, request a prediction length, and get both point forecasts and quantile forecasts back. That makes it one of the more interesting non-transformer baselines in the hosted catalog, especially if you want to compare architectural families rather than just checkpoint sizes.

#Architecture: xLSTM Instead of a Transformer

The TiRex paper, TiRex: Zero-Shot Forecasting Across Long and Short Horizons with Enhanced In-Context Learning, argues that recurrent state tracking remains valuable for forecasting even as transformers dominate neighboring domains. The authors position xLSTM as a way to combine stronger in-context learning behavior than classic LSTMs with state-tracking properties that are useful for long-horizon time-series problems.

That framing matters because many third-party summaries initially described TiRex as just another transformer. It is not. The official model card describes TiRex as a 35M-parameter xLSTM forecasting model, and the paper explicitly contrasts it with transformers, state-space models, and RWKV-style architectures rather than grouping it with them.

The paper also introduces a training-time masking strategy called CPM, which is meant to improve in-context forecasting behavior. If you are evaluating TiRex, that combination of xLSTM backbone plus masking strategy is the architectural story to pay attention to, not any transformer-style encoder-decoder interpretation.

#What the Current Public Release Supports

The official Hugging Face quick start is narrow and concrete: load NX-AI/TiRex, pass a context tensor, and call forecast(..., prediction_length=...). The model card highlights three core properties:

• zero-shot forecasting,
• point plus quantile predictions,
• strong benchmark performance across short and long horizons.

Just as important is what the public release does not currently document. The official materials do not expose a covariate API for the released forecasting model, and NX-AI clarified in a Hugging Face discussion that the current version of TiRex does not support covariates, though future versions may. The same discussion also confirms that the current model is univariate rather than a general multivariate forecasting surface.

So if you need external regressors, holidays, promotions, weather, or similar side inputs today, TiRex should not be the default assumption. Choose a model whose official card explicitly documents covariate handling.

#Training Data and Benchmarks

The official model card lists autogluon/chronos_datasets and Salesforce/GiftEvalPretrain as the named training datasets for the released checkpoint. The paper then evaluates TiRex on benchmark suites including GiftEval and Chronos-ZS, where the authors report state-of-the-art zero-shot forecasting performance despite TiRex being much smaller than several transformer competitors.

That benchmark framing is consistent with how the current public checkpoint should be understood: a compact, benchmark-oriented zero-shot forecaster rather than a broad "everything model" that already covers covariates, multivariate reasoning, and rich control inputs.

#When to Choose TiRex

TiRex makes the most sense when you want:

• a compact non-transformer forecasting baseline,
• zero-shot forecasting with probabilistic outputs,
• a model that is explicitly optimized for both short- and long-horizon evaluation.

It is a weaker fit when you already know you need covariates, explicit multivariate input structure, or a broader documented production ecosystem. In those cases, the constraint is not that TiRex is low quality; it is that the current public release is narrower than some third-party summaries suggested.

#Availability on TSFM.ai

TiRex is available in the TSFM.ai model catalog through the NX-AI/TiRex entry, and the catalog also includes NX-AI/TiRex-1.1-gifteval as a related variant. The hosted catalog description should be read in line with the official model card: xLSTM-based, 35M parameters, zero-shot forecasting, and point plus quantile outputs.

If you are evaluating TiRex on TSFM.ai, use it as a high-quality univariate zero-shot forecaster. If you need richer input structure, treat that as a model-selection question rather than assuming TiRex already supports it.