NX-AI

NXAI builds efficient foundation models on xLSTM, a recurrent architecture that scales linearly in sequence length rather than quadratically like Transformer attention.

Our flagship, TiRex 🦖, is a zero-shot time series forecasting model deployed in industrial and edge settings.

⚙️ What We Build

• xLSTM: a recurrent architecture led by Sepp Hochreiter (co-inventor of LSTM, 1997). Its recurrence scales linearly in sequence length, versus the quadratic cost of self-attention, which is the basis for its efficiency claims on long-sequence and state-tracking tasks.
• Foundation models for industrial data: time series, vision and multimodal.
• Edge-ready inference: models sized and quantised for embedded and cloud-edge deployment, not data-centre-only inference.
• TiRex 🦖: a 35M-parameter zero-shot forecasting foundation model built on xLSTM, for industrial time series.
• TiRex-2 🦖: our current flagship forecasting model (38.4M active parameters, +44.1M for multivariate mode), extending TiRex with native multivariate forecasting and past/future-known covariates.

📚 Publications & Resources

All papers, code and live demos for xLSTM, TiRex, Vision-LSTM, Bio-xLSTM and LRAM in one place: nx-ai.github.io/linktree

🔬 Collaborate With Us

We welcome:

• Industrial and enterprise partners deploying efficient AI at production scale
• Researchers and developers working on xLSTM and efficient foundation models
• Open-source contributions: fork our repositories, file issues, submit pull requests