Technical Report on Ling and Ring 2.6: Trillion-Parameter-Scale Instant Agent Intelligence

This article introduces the technical report on the Ling-2.6 and Ring-2.6 model families released by the arXiv team (original link: http://arxiv.org/abs/2606.15079v1, published on 2026-06-13). The report proposes to achieve efficient and scalable agent intelligence at the trillion-parameter scale through collaborative architecture design, hybrid linear attention, evolutionary chain-of-thought, and the KPop reinforcement learning framework. The two models have clear divisions of labor: Ling-2.6 focuses on instant responses, while Ring-2.6 excels at deep reasoning; all checkpoints are open-sourced to provide resources for the community.

Efficient and scalable agent intelligence needs to meet three conflicting requirements: low-latency response, strong reasoning ability, and trainable deployment. Existing models struggle to balance these. Although trillion-parameter models are powerful, their training and inference costs are high, and demands such as long-context processing and multi-step reasoning further intensify the tension between efficiency and performance.

Ling-2.6 is an expert in instant responses, suitable for scenarios like real-time dialogue and fast code completion; Ring-2.6 is an expert in deep reasoning, applicable to tasks such as complex planning and tool coordination. The division-of-labor strategy avoids resource waste and dynamically matches task characteristics.

1. Architecture migration pre-training: Upgraded based on Ling-2.0 to reduce training costs; 2. Unified collaborative design: Balances model architecture, optimization objectives, service systems, and agent training environments; 3. Hybrid linear attention: Integrates Lightning Attention and MLA to improve long-context training and decoding efficiency; 4. Token efficiency optimization: Evolutionary chain-of-thought (compact reasoning), language unit strategy optimization (reinforcement learning to select effective tokens), bidirectional preference alignment (balancing efficiency and human satisfaction), and shortest correct response distillation.

The KPop framework, designed for Ring-2.6-1T, supports training with large-scale environmental data; its asynchronous scheduling mechanism can handle tasks such as programming, search, tool usage, and workflow execution; through agent-environment interaction learning, it cultivates complex agent capabilities.

Training efficiency: Architecture migration reduces computing resources, and collaborative design narrows the gap between training and deployment; Inference efficiency: Hybrid attention + token optimization reduces latency and improves throughput; Agent capabilities: The KPop framework supports diverse tasks and performs excellently.

All checkpoints are open-sourced, with values including lowering research barriers, promoting transparent research, accelerating application development, and driving standardization; Application scenarios: Ling is suitable for real-time interaction systems (chatbots, real-time assistants), Ring is suitable for complex task automation (research analysis, code generation), and the two collaborate to build hybrid agent systems.

Ling-2.6 and Ring-2.6 are important steps toward practical agent intelligence, achieving the unification of efficiency and capability at the trillion-parameter scale. Technical insights: The importance of collaborative design, balance between specialization and generalization, unification of efficiency and capability, and potential of environment-based learning. We will continue to promote the development of agent technology in the future.