Ant Bailing Ring-2.6-1T Open Source with Enhanced Agent Execution Capabilities
Ant Group's Bailing releases open-source reasoning model Ring-2.6-1T, featuring on-demand reasoning with adjustable effort, scoring 95.83 on AIME 26, and trained with asynchronous reinforcement learning.
On May 15, Ant Group’s Bailing announced the open-source release of its flagship reasoning model Ring-2.6-1T, with weights available on Hugging Face and ModelScope. The model had previously been launched on OpenRouter with a limited free API trial.
Ring-2.6-1T is designed with an “on-demand thinking” philosophy, featuring an adjustable Reasoning Effort mechanism with two modes: high and xhigh. High mode targets frequent agent workflows for higher efficiency, suitable for multi-turn dialogues, tool collaboration, and task decomposition. Xhigh mode is geared toward high-difficulty tasks such as math competitions and scientific analysis, unlocking the model’s full potential. Developers have praised this as a practical engineering advancement.
Benchmark results show the model excels in both modes. In high mode, it scored 87.60 on PinchBench, outperforming GPT-5.4 xHigh and Gemini-3.1-Pro high, and 95.32 on Tau2-Bench Telecom, demonstrating significant agent execution capabilities. In xhigh mode, it achieved 95.83 on AIME 26, approaching top-tier models, and 88.27 on GPQA Diamond, indicating robust scientific knowledge understanding and complex reasoning.
Training utilizes an asynchronous (Async) reinforcement learning architecture that decouples policy sampling from parameter updates into independent pipelines, overcoming the GPU idle time and low throughput issues of traditional synchronous training. This architecture supports longer continuous training periods. Bailing also incorporated the “popsicle algorithm” previously validated in Ring-1T into the Async RL framework to stabilize training. Detailed technical specifications will be disclosed in a forthcoming report.
Over the past month, Bailing has iterated and open-sourced multiple models, including the Ling language model series and Ring reasoning series. Notably, instead of focusing on larger parameter counts or higher single-point scores, Bailing emphasizes real-world production usage, with the entire model series prioritizing “Token Efficiency” — achieving high-quality outputs with fewer tokens. Market feedback has been positive: the anonymous test version of Ling-2.6-flash, codenamed “Elephant Alpha,” ranked first on OpenRouter’s trending list for consecutive days, with daily token calls reaching 100B level.