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

TL;DR

A model-family report on Ling-2.6 and Ring-2.6, combining architectural migration, long-context efficiency work, token-efficient reasoning objectives, and asynchronous agent reinforcement learning at trillion-parameter scale.

Why it matters

The report connects model architecture, serving efficiency, post-training, and environment throughput. Its KPop framework is explicitly designed to keep large-scale agent RL stable across coding, search, tool use, and workflow environments.

Key findings

1. 01

   Ring-2.6-1T is trained with KPop, an RL framework built around large-scale environment-grounded data.

2. 02

   KPop uses asynchronous scheduling across coding, search, tool-use, and workflow execution to improve training efficiency.

3. 03

   The broader recipe also optimizes capability per output token, making response length and serving cost first-class scaling variables.

Scaling dimensions

Models, methods & benchmarks

Models

Ling-2.6, Ring-2.6, Ring-2.6-1T

Algorithms

KPop, Reinforcement learning, Evolutionary Chain-of-Thought, Linguistic Unit Policy Optimization

Topics