Ring-2.5-1T - Ant Financial's open-source trillion-parameter thinking model

What is Ring-2.5-1T?

Ring-2.5-1T is a trillion-parameter hybrid linear thinking model open-sourced by Ant Group's inclusionAI team. It adopts the Ling 2.5 architecture and integrates MLA and Lightning Linear attention mechanisms in a 1:7 ratio. The model achieves a 10x reduction in memory access and a 3x increase in throughput when generating text longer than 32K. It achieved gold medal levels in the IMO 2025 and CMO 2025 mathematics competitions, natively adapts to intelligent agent frameworks such as Claude Code, supports multi-step planning and tool calls, and is suitable for deep reasoning and long-term task execution scenarios.

Main functions of Ring-2.5-1T

• Efficient generation of long textsIn scenarios with sequence lengths of 32K or more, the hybrid linear attention architecture reduces memory access overhead to one-tenth of that of traditional models while achieving more than three times the generation throughput improvement.
• Mathematical Deep ReasoningBased on intensive reward reinforcement learning training, the student achieved scores of 35 and 105 in the International Mathematical Olympiad (IMO) 2025 and the China Mathematical Olympiad (CMO) 2025 respectively, both exceeding the gold medal score threshold.
• Intelligent agent task executionIt natively adapts to intelligent agent frameworks such as Claude Code and OpenClaw, supporting multi-step planning, tool calls, and autonomous completion of complex software engineering tasks for several consecutive hours.

Technical Principles of Ring-2.5-1T

• Hybrid linear attention architectureRing-2.5-1T transforms the traditional GQA layer into a combined structure of MLA and Lightning Linear at a ratio of 1:7. The Lightning Linear layer processes long sequence generation with linear complexity, while the MLA layer further compresses the KV cache and enhances expressive power.
• Intensive Reward Reinforcement LearningThe model is trained using a reinforcement learning framework based on verifiable rewards, and a dense reward mechanism is introduced to provide feedback on the rigor of the reasoning process, enabling the model to generate logically rigorous and complete mathematical proofs in a deep thinking mode.
• Asynchronous agent reinforcement learningFor long-range intelligent agent tasks, Ring-2.5-1T improves autonomous decision-making capabilities through large-scale fully asynchronous reinforcement learning training, enabling it to continuously execute multi-step operations without interruption in complex environments.

Project address for Ring-2.5-1T

• GitHub repository: https://github.com/inclusionAI/Ring-V2.5
• HuggingFace model libraryhttps://huggingface.co/inclusionAI/Ring-2.5-1T

Application scenarios of Ring-2.5-1T

• Long document processing and generationIt is suitable for professional writing scenarios that require processing tens of thousands of words of context, such as academic paper writing, technical document generation, and novel writing, and significantly reduces the computational cost and response latency of long text generation.
• Mathematical Research and Educational SupportAs an intelligent coach for mathematics competition training, it provides IMO/CMO level problem-solving approaches and complete proof processes, or assists researchers in exploring the formal verification of complex mathematical conjectures.
• Independent Software EngineeringThe model can independently complete long-term programming tasks such as operating system development, large codebase refactoring, and multi-file collaborative programming, and achieve end-to-end software delivery in intelligent agent frameworks such as Claude Code.
• Complex Task Planning and ExecutionIt is suitable for intelligent agent application scenarios that require cross-tool collaboration, such as in-depth research, data analysis workflow construction, and multi-step business decision support, and supports autonomous task execution for several hours.