🔔 Ant Group Launches Trillion-Parameter Reasoning Model Ring-2.6-1T: PinchBench Score of 87.60 Surpasses GPT-5.4 According to Dongcha Beating Monitoring, Ant Group’s Bailing Large Model Team has launched its flagship trillion-parameter reasoning model, Ring-2.6-1T (with 63 billion activated parameters). Designed specifically for complex tasks and production environments, the model introduces a new “Dynamic Reasoning Intensity” mechanism that enables flexible balancing between cognitive depth, token cost, and execution speed. Based on varying computational demands, the model offers two operational modes: high and xhigh. In its Agent mode (high), optimized for multi-step execution and tool usage, it achieves a PinchBench score of 87.60, outperforming GPT-5.4 xHigh and Gemini-3.1-Pro high, with a ClawEval score of 63.82. In its deep reasoning mode (xhigh), tailored for mathematical reasoning and scientific research, it scores 95.83 on AIME 26 and 88.27 on GPQA Diamond. The official team notes that text formatting and math competition tasks demand vastly different computational resources; this mechanism is designed to reduce token overhead, enabling the model to serve as the default foundation for high-frequency scenarios such as tool orchestration, programming, and multi-turn interactions. Starting today, in collaboration with Novita, a free API trial is available on the OpenRouter platform until May 15, and model weights will be open-sourced soon. https://t.co/M2jDnq3iis