TiRex 1.1 (GiftEval)

TiRex 1.1 GiftEval is NX-AI's updated xLSTM-based zero-shot forecasting checkpoint with the TiRex 1.1 inference improvements and cleaned pretraining data to remove GIFT-Eval test overlap. It preserves the compact non-Transformer architecture of the original TiRex while tightening the evaluation setup for the published benchmark release, so it is the variant to reach for when leakage-clean benchmark comparability matters alongside forecast quality.

Like the base TiRex it is a 35M-parameter xLSTM forecasting model rather than a Transformer, and it produces both point and quantile forecasts across short and long horizons. Its provenance is documented more precisely than the base checkpoint: the official card lists filtered autogluon/chronos_datasets, a GiftEvalPretrain subset, and synthetic data, and NX-AI notes that the 1.1 release additionally applies long period normalisation at inference time.

On TSFM.ai, pick this checkpoint when you want the latest TiRex inference behaviour and a pretraining mixture explicitly scrubbed of GIFT-Eval overlap. Choose the base NX-AI/TiRex instead when you simply want the original compact xLSTM forecaster and do not need the cleaned-data, benchmark-oriented configuration. Both share the same architecture and parameter count, so the choice is about evaluation hygiene and inference updates rather than capacity.