InclusionAI/Ring-2.6-1T is now open-sourced (huggingface.co)

InclusionAI has announced the open-source release of Ring-2.6-1T, a trillion-parameter reasoning model tailored for complex real-world tasks across various sectors such as engineering, scientific research, and enterprise automation. Unlike previous models that focused primarily on scaling parameters, Ring-2.6-1T enhances execution capabilities, enabling it to perform multi-step tasks, plan actions, and collaborate with tools effectively. This model introduces significant advancements in agent execution with its adjustable Reasoning Effort mechanism, allowing developers to choose between "high" and "xhigh" reasoning levels depending on task complexity, significantly optimizing performance and resource use for different workflows. Key innovations include an asynchronous reinforcement learning (RL) training paradigm supported by the IcePop algorithm, which improves training stability and efficiency for trillion-parameter architectures. This allows for better resource utilization during training and supports longer, uninterrupted training cycles. Early benchmark results show Ring-2.6-1T outperforming competitors like GPT-5.4 and Gemini-3.1-Pro in real-world task execution scenarios. Overall, this development positions Ring-2.6-1T as a versatile tool that can understand user intent and drive complex, iterative workflows, promising significant implications for the AI/ML community in automation and advanced reasoning tasks.