Ant launches trillion-parameter thinking model Ring-2.6-1T: PinchBench scores 87.60, surpassing GPT-5.4

According to Beating Monitoring, Ant Group’s Bailing Large Model Team has launched the trillion-parameter flagship reasoning model Ring-2.6-1T (active parameters 63 billion). This model is specifically designed for complex tasks and production environments. Its core novelty is a “dynamic thinking intensity” mechanism, enabling the system to flexibly balance between cognitive depth, Token cost, and execution speed.

Based on different compute load requirements, the model provides two operating modes: high and xhigh. In the Agent mode (high), which focuses on multi-step execution and tool calls, its PinchBench score reaches 87.60, higher than GPT-5.4 xHigh and Gemini-3.1-Pro high, and its ClawEval test score is 63.82. In the deep-thinking mode (xhigh) for mathematical reasoning and scientific research, its AIME 26 score is 95.83, and its GPQA Diamond score is 88.27.

Officially, text-format conversion and math competitions have markedly different demands on compute power. The purpose of designing this mechanism is to reduce Token overhead, allowing the model to serve as the default foundation for high-frequency scenarios such as tool orchestration, programming, and multi-turn interactions. Starting today, the model will jointly with Novita offer a one-week free API trial on the OpenRouter platform (until May 15), and its weights will be open-sourced soon.