TiRex
35M-parameter xLSTM-based zero-shot forecaster. Uses Contiguous Patch Masking during training to stabilise long-horizon autoregressive generation. Reports SOTA on GIFT-Eval and Chronos-ZS, outperforming much larger transformer models.
TiRex is NX-AI's argument that recurrence isn't dead. Where most TSFMs are transformers (or, more recently, state-space models), TiRex uses xLSTM, an enhanced LSTM variant from the Hochreiter group. The claim: xLSTM retains genuine state-tracking, which transformers, SSMs, and parallelisable RNNs like RWKV all struggle with — and state-tracking matters most for long-horizon forecasting.
The training recipe contributes Contiguous Patch Masking(CPM), a masking strategy that improves coherence in multi-step autoregressive generation by mitigating the usual error accumulation pattern. At 35M parameters, TiRex reports SOTA results on GIFT-Eval and Chronos-ZS, beating Chronos-Bolt, TimesFM, and Moirai head-to-head at much smaller cost.
Versions on TS-Arena
Each version below corresponds to one registered model id in the leaderboard. Click through to its detail page for per-model rankings, forecasts, and history.
- TiRextirex35M params…
Single released checkpoint; xLSTM-based.