Ant Group Open Sources Ring-2.5-1T, World's First Trillion-Parameter Thinking Model with Mixture of Linear Architecture

On February 13, Ant Group made the world's first trillion-parameter reasoning model, Ring-2.5-1T, based on a hybrid linear architecture, available as open source. It delivers state-of-the-art open-source performance in long-text generation, mathematical reasoning, and agent task execution, offering a high-performance foundation for handling complex tasks in the agent era.

For generation efficiency, in long-text scenarios beyond 32K, Ring-2.5-1T's memory access is over 10 times lower than the previous generation, while throughput is more than three times higher. In deep reasoning, the model achieved gold-medal-level scores (35 for IMO, 105 for CMO) in self-testing for the International Mathematical Olympiad (IMO2025) and the Chinese Mathematical Olympiad (CMO2025). It also integrates seamlessly with agent frameworks such as Claude Code and OpenClaw personal AI assistants, enabling multi-step planning and tool calling.

(Figure caption: Ring-2.5-1T achieves open-source leading performance in high-difficulty reasoning tasks like mathematics, code, and logic, as well as in long-term execution tasks such as agent search, software engineering, and tool calling.)

Across multiple authoritative benchmarks, Ring-2.5-1T was systematically compared with leading open-source and closed-source models, including 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 performance in high-difficulty scenarios like mathematical reasoning, code generation, logical reasoning, and agent task execution. Particularly in Heavy Thinking mode, it outperformed all compared models on mathematical competition benchmarks such as IMOAnswerBench and HMMT-25, as well as on the LiveCodeBench-v6 code generation benchmark, demonstrating strong complex reasoning and cross-task generalization.

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 number of activated parameters increased from 51B in the previous generation to 63B. Supported by the hybrid linear attention architecture, the reasoning efficiency is significantly improved over the previous generation. Compared to the KIMI K2 architecture, which has only 32B activated parameters, the Ling2.5 architecture demonstrates significant throughput advantages for long-sequence reasoning tasks at a total parameter count of 1T, and this efficiency advantage grows with increasing generation length.

(Figure caption: Efficiency comparison across different generation lengths. The longer the generation, the more pronounced 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 addresses the challenges of high computational costs and slow reasoning speed in long output scenarios. The open-sourcing of this model also highlights the Ant Bailing team's comprehensive capabilities in large-scale training infrastructure, algorithm optimization, and engineering implementation, offering the industry a new choice for high-performance, high-efficiency foundation models in the agent era.

Currently, the model weights and inference code for Ring-2.5-1T have been released on major open-source platforms like Hugging Face and ModelScope. The official platform's Chat experience page and API service will be available soon.