On February 13, Ant Group open-sourced the world's first trillion-parameter reasoning model based on a hybrid linear architecture, Ring-2.5-1T. It achieves leading open-source performance in long-text generation, mathematical reasoning, and agent task execution, providing high-performance foundational support for complex task processing in the era of agents.
In terms of generation efficiency, in long-text generation scenarios above 32K, the memory access of Ring-2.5-1T is reduced by more than 10 times compared to the previous generation model, and the throughput is improved by more than 3 times. In deep thinking capabilities, the model has achieved gold medal levels (35 points for IMO, 105 points for CMO) in self-testing for the International Mathematical Olympiad (IMO2025) and the Chinese Mathematical Olympiad (CMO2025). At the same time, it can easily adapt to agent frameworks like Claude Code and OpenClaw personal AI assistants, supporting multi-step planning and tool calling.
(Figure caption: Ring-2.5-1T reaches open-source leading levels in high-difficulty reasoning tasks such as mathematics, code, and logic, as well as in long-term task execution such as agent search, software engineering, and tool calling.)
In multiple authoritative benchmark tests, Ring-2.5-1T was systematically compared with mainstream open-source and closed-source models such as DeepSeek-v3.2-Thinking, Kimi-K2.5-Thinking, GPT-5.2-thinking-high, Gemini-3.0-Pro-preview-thinking, and Claude-Opus-4.5-Extended-Thinking. It achieved open-source leading levels in high-difficulty scenarios such as mathematical reasoning, code generation, logical reasoning, and agent task execution. Especially in the Heavy Thinking mode, the model surpassed all comparison models in mathematical competition reasoning benchmarks like IMOAnswerBench and HMMT-25, and in the LiveCodeBench-v6 code generation benchmark, demonstrating strong complex reasoning and cross-task generalization capabilities.
Ring-2.5-1T is based on the Ling2.5 architecture. By optimizing the attention mechanism, it significantly improves the efficiency and stability of long-text reasoning. The activated parameter scale increased from 51B in the previous generation to 63B. However, supported by the hybrid linear attention architecture, the reasoning efficiency has been significantly improved compared to the previous generation. Compared with the KIMI K2 architecture that only has 32B activated parameters, the Ling2.5 architecture still shows significant advantages in throughput for long-sequence reasoning tasks at a total parameter count of 1T, and the efficiency advantage continues to expand as the generation length increases.
(Figure caption: Efficiency comparison under different generation lengths. The longer the generation length, the more obvious the throughput advantage.)
As AI large model applications expand from short conversations to long document processing, cross-file code understanding, and complex task planning, Ring-2.5-1T effectively alleviates the issues of high computational costs and slow reasoning speed in long output scenarios. The open sourcing of this model also reflects the comprehensive capabilities of the Ant Bailing team in large-scale training infrastructure, algorithm optimization, and engineering implementation, providing the industry with a new choice of high-performance, high-efficiency foundational models for the era of agents.
Currently, the model weights and inference code of Ring-2.5-1T have been released on mainstream open-source platforms such as Hugging Face and ModelScope. The official platform's Chat experience page and API service will be launched shortly.
