Ling and Ring 2.6 Technical Report: Efficient and Instant Agentic Intelligence at Trillion-Parameter Scale

The technical report introduces Ling-2.6 and Ring-2.6, models designed for efficient, scalable agentic intelligence at a trillion-parameter scale. Ling-2.6 focuses on low-latency response generation with enhanced output token efficiency, while Ring-2.6 is optimized for advanced reasoning and agentic workflows, utilizing a hybrid linear attention architecture and a reinforcement learning framework called KPop for improved training on complex tasks. The open-sourcing of these models aims to facilitate further research and development in practical agentic systems, addressing the challenges of training and deploying large-scale intelligent agents.