Ling-2.6

Ant Group's trillion-parameter non-thinking flagship (backfilled; weights April 23, technical report June 13). Hybrid Multi-head Latent Attention + linear attention at 1T scale, MIT-licensed. The joint Ling & Ring 2.6 technical report ("Efficient and Instant Agentic Intelligence at Trillion-Parameter Scale") covers both this model and the reasoning sibling Ring-2.6: "fast thinking" trained via a Contextual Process Redundancy Suppression reward, plus async RL with IcePop for trillion-parameter RL stability.