TiRex-2

TiRex-2 is NX-AI's second-generation xLSTM zero-shot forecaster and a substantial upgrade over the original TiRex. It keeps the compact, recurrent, non-Transformer backbone that distinguishes the TiRex line, but extends it from univariate point-and-quantile forecasting to native multivariate forecasting and covariate conditioning, so correlated series can be forecast jointly and known drivers can be fed in alongside the target. Like every TiRex checkpoint it emits a full grid of quantile forecasts directly, giving calibrated probabilistic output across short and long horizons without any task-specific training.

Architecturally it is an sLSTM/mLSTM (xLSTM) model that activates roughly 38.4M parameters for univariate forecasting and an additional 44.1M when it engages its variate-mixing path for multivariate targets. That recurrent design gives it in-context generalization behaviour unlike the attention-only models that dominate this catalog, and the multivariate and covariate paths are the headline capabilities that set TiRex-2 apart from the 35M-parameter original.