Sustainable, Low-Energy, and Fast AI Made in Europe

Description

The rapid scaling of large AI models has exposed inherent limitations in attention-based architectures, whose heavy computational and energy demands restrict deployment in industrial, automotive, and edge environments. We introduce a new class of sustainable, low-energy, and fast AI systems built on the xLSTM architecture and its time-series foundation model, TiRex. xLSTM employs a memory-centric, sub-quadratic design that matches transformer-level reasoning performance while dramatically reducing compute load, latency, and energy consumption. TiRex extends these advantages to temporal data, delivering state-of-the-art results across forecasting, anomaly detection, and control tasks at a fraction of the operational cost of existing models. Crucially, TiRex leverages xLSTM’s intrinsic state-tracking capabilities that are absent in Transformer and State-Space architectures, thereby enabling stable long-horizon forecasting and superior generalization. Together, xLSTM and TiRex define a new paradigm of efficient intelligence: models that are not only faster and more robust, but intrinsically aligned with the sustainability and efficiency demands of modern industrial AI. This architecture supports real-time on-device inference, large-scale distributed deployment, and a pathway toward environmentally responsible AI innovation without sacrificing accuracy or capability.

Period	04 Dec 2025
Event title	EurIPS 2025
Event type	Conference
Location	DenmarkShow on map
Degree of Recognition	International