The Ling-2.6-1T, a trillion-parameter flagship model developed by the Bailing Large Model team under Ant Group, was officially announced today to be open-sourced for developers. Rather than blindly pursuing the stacking of parameter scale, the model is dedicated to systematically optimizing the model's instruction execution, tool adaptation, and long-context handling capabilities in real and complex tasks.

image.png

Architectural Innovation Enables Efficient "Fast Thinking" Mechanism

Ling-2.6-1T adopts an innovative hybrid architecture, successfully reducing token costs through an enhanced reward strategy that suppresses process redundancy. This "fast thinking" mechanism allows the model to output high-quality results at a lower cost while maintaining the upper limit of trillion-parameter intelligence, significantly improving the intelligence efficiency ratio.

For complex workflows, the model enhances learning of composite tasks. In multiple authoritative benchmark evaluations for execution tasks, Ling-2.6-1T demonstrates strong multi-step execution capabilities, reaching the top level in the open-source field in code generation, defect fixing, and precise reasoning in noisy environments.

Full-Stack Compatibility Helps Enterprise Production Workflows Go Live

To bring the trillion-parameter model into developers' practical production, Ling-2.6-1T achieves high compatibility with mainstream Agent frameworks. It can adapt to complex business scenarios with multiple tools and constraints, aiming to become a core capability foundation that can be deployed and continuously operated within enterprise systems.

Currently, the model has been launched on open-source communities such as Hugging Face and ModelScope. To support global developers' experience and evaluation, the official also announced an extension of one week for the free API call service on the OpenRouter platform, further lowering the entry barrier for using the trillion-parameter model.

  • Hugging Face: https://huggingface.co/inclusionAI/Ling-2.6-1T

  • ModelScope: https://www.modelscope.cn/models/inclusionAI/Ling-2.6-1T