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

Ling-2.6 and Ring-2.6 are MoE models at trillion-parameter scale built by upgrading the Ling-2.0 base via architectural migration pre-training rather than training from scratch. The key architectural addition is a hybrid linear attention mechanism aimed at reducing inference latency while preserving reasoning quality. Ring-2.6 targets multi-step agentic workflows while Ling-2.6 prioritizes low-latency single-turn response generation.