Ant open source trillion-parameter thinking model Ring-2.5-1T

On February 13th, Ant Group released and open-sourced the first trillion-parameter hybrid linear architecture thinking model Ring-2.5-1T. Currently, the model weights and inference code for Ring-2.5-1T have been released on mainstream open-source platforms such as Hugging Face and ModelScope. The official platform Chat experience page and API services will be launched soon. As a crucial step towards the era of general intelligence, the team has significantly expanded the hybrid linear attention architecture in both pre-training and reinforcement learning. On one hand, they utilize the efficient 1:7 MLA + Lightning Linear Attention architecture to enhance the model's thinking efficiency and exploration space. On the other hand, they expand reinforcement learning and intelligent agent environment scale to improve the model's thinking depth and long-range execution capability. Compared to the previously released Ring-1T, Ring-2.5-1T has made significant improvements in generation efficiency, thinking depth, and long-range execution: Efficient generation: Thanks to the high proportion of linear attention mechanism, the memory access scale decreases by more than 10 times and generation throughput increases by more than 3 times for generation lengths exceeding 32K, making it particularly suitable for tasks requiring deep thinking and long-range execution. Deep thinking: Introducing dense reward based on RLVR to provide feedback on the rigor of the thinking process, Ring-2.5-1T achieves gold medal level in both IMO 2025 and CMO 2025 (self-tested). Long-range execution: Through large-scale fully-async agentic RL training, the model significantly improves its long-range autonomous execution capability for complex tasks, enabling Ring-2.5-1T to easily adapt to intelligent agent programming frameworks like Claude Code and personal AI assistant OpenClaw.