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 fuses MLA with Lightning Linear attention mechanisms in a 1:7 ratio. When generating long texts above 32K, the model reduces memory access by 10x and increases throughput by 3x. It reaches gold medal level in the IMO 2025 and CMO 2025 mathematics competitions, and natively adapts toClaude Codeand other agent frameworks, supporting multi-step planning and tool calling, and is suitable for deep reasoning and long-horizon task execution scenarios.
Main Features of Ring-2.5-1T
-
Efficient Long-Text Generation: In scenarios with sequence lengths above 32K, the hybrid linear attention architecture reduces memory access overhead to one tenth that of traditional models while achieving more than three times the generation throughput.
-
Deep Mathematical Reasoning: Trained with dense-reward reinforcement learning, it achieved 35 points and 105 points in the International Mathematical Olympiad IMO 2025 and the Chinese Mathematical Olympiad CMO 2025 respectively, both exceeding the gold medal cutoff.
-
Agent Task Execution: Natively adapts to Claude Code andOpenClawand other agent frameworks, supporting multi-step planning, tool calling, and autonomous completion of complex software engineering tasks over several consecutive hours.
Technical Principles of Ring-2.5-1T
- Hybrid Linear Attention Architecture: Ring-2.5-1T transforms traditional GQA layers into a combined structure of MLA and Lightning Linear in a 1:7 ratio. The Lightning Linear layers process long-sequence generation with linear complexity, while the MLA layers further compress the KV cache and enhance expressiveness.
- Dense-Reward Reinforcement Learning: The model is trained using a reinforcement learning framework based on verifiable rewards, introducing a dense reward mechanism to provide feedback on the rigor of the reasoning process, enabling the model to generate mathematically rigorous proofs with complete steps in deep thinking mode.
- Asynchronous Agent Reinforcement Learning: For long-horizon agent tasks, Ring-2.5-1T improves autonomous decision-making through large-scale fully asynchronous reinforcement learning training, enabling it to continuously execute multi-step operations in complex environments without interruption.
Project Addresses of Ring-2.5-1T
- GitHub Repository：https://github.com/inclusionAI/Ring-V2.5
- HuggingFace Model Library：https://huggingface.co/inclusionAI/Ring-2.5-1T
Application Scenarios of Ring-2.5-1T
-
Long Document Processing and Generation: Suitable for professional writing scenarios such as academic paper writing, technical document generation, and long novel creation that require handling tens of thousands of words of context, greatly reducing the computational cost and response latency of long-text generation.
-
Mathematical Research and Education Assistance: As an intelligent coach for mathematics competition training, it provides IMO/CMO-level problem-solving approaches and complete proof processes, or assists researchers in exploring formal verification of complex mathematical conjectures.
-
Autonomous Software Engineering: The model can independently complete long-horizon programming tasks such as operating system development, large-scale codebase refactoring, and multi-file collaborative programming, achieving end-to-end software delivery in agent frameworks such as Claude Code.
-
Complex Task Planning and Execution: Suitable for agent application scenarios such as deep research requiring cross-tool collaboration, data analysis workflow construction, and multi-step business decision support, supporting autonomous task execution lasting several hours.