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

Summary

Inclusion AI Ling-2.6 (instant) and Ring-2.6 (thinking) upgrade Ling-2.0 via hybrid Lightning Attention+MLA (7:1), ~9.6T continue-pretrain tokens, and native agentic RL. KPop replaces IcePop’s fixed-ratio double-sided mask with a symmetric binary-KL token mask (one hyperparameter φ), adapting to train–inference mismatch on long-horizon rollouts. Paper: Ring-2.6-1T SWE-bench Verified 70.8%→76.28% with KPop; tweet claims >76 with pure RL. Open-sourced Ling-2.6-flash, Ling-2.6-1T, Ring-2.6-1T.

Keywords

kpop · icepop · ring-2.6 · ling-2.6 · moe · agentic-rl · inclusionai · ant-group · x

Topics

agentic LLMs, MoE RL, hybrid linear attention

Research notes

• Primary: arxiv abs 2606.15079. Discord/X https://x.com/Jia__Guo/status/2059291333496553797 via fxtwitter announcing KPop. Notion blog https://ringtech.notion.site/kpop (JS-gated at fetch). Weights HF collections inclusionAI/ling-26 and inclusionAI/ring-26. Code https://github.com/inclusionAI/Ling-V2.5 (MIT). Corresponding Jun Zhou / Zhiqiang Zhang. Open weights, not a new hosted corpus, so no datasets_local row.