Overview
Ant Group's Bailing Ring-2.6-1T is a trillion-parameter flagship reasoning model officially released by Ant Group's inclusionAI team on May 9, 2026. Based on a trillion-level MoE architecture, it targets real-world complex tasks such as Agent, engineering development, and scientific research analysis, featuring adjustable reasoning intensity and efficient execution. It scored 87.60 on the PinchBench test, surpassing multiple international mainstream closed-source models, and was open-sourced on May 15, 2026, with weights on Hugging Face and ModelScope.
Key Features
- Trillion-parameter MoE architecture: 1T parameter scale, MoE activation mechanism controls reasoning cost while maintaining high parameters
- Adjustable reasoning intensity: Supports reasoning modes like high/xhigh, flexibly adjusting thinking depth based on task complexity
- PinchBench leading: 87.60 points on PinchBench test, surpassing international closed-source models like GPT
- Agent scenario optimization: Designed for Agent workflows, supporting complex task decomposition and execution
- Open-sourced: Weights available on Hugging Face and ModelScope
Use Cases
- Complex task execution for AI Agents
- Software engineering development assistance
- Scientific research data analysis and paper comprehension
- Math and programming competitions
- Enterprise-level reasoning-intensive applications
Pros
- Trillion-parameter scale, one of the largest open-source models in China
- Adjustable reasoning intensity, flexibly adapting to different scenarios
- PinchBench surpasses international closed-source models
- Open source + limited-time free, developer-friendly
Pricing
Limited-time free API calls will be available soon. Model weights are open-sourced for self-deployment.
Summary
Ant Group's Bailing Ring is a representative domestic trillion-parameter reasoning model, excelling in reasoning-intensive tasks. Its open-source release makes it a strong choice for developers building Agent applications.
Version History
- Ring-2.6-1T officially open-sourced (2026-05-15): Weights released on Hugging Face and ModelScope, with high/xhigh reasoning-intensity modes