NVIDIA AI Unveils Gated DeltaNet-2 for Enhanced Model Efficiency

NVIDIA has introduced Gated DeltaNet-2, a new linear attention layer designed to improve the efficiency of AI models. This innovation aims to address bottlenecks in compressed memory editing by decoupling erase and write operations. The model, trained with 1.3 billion parameters, reportedly outperforms previous versions and similar models across various research benchmarks.