A Chinese company has released 'Ring-2.6-1T,' an AI model worth 1 trillion tokens, as an open model, surpassing GPT-5.4 and Gemini 3.1 Pro in some benchmark tests.



InclusionAI , an AI development team operating under the umbrella of the Chinese financial company Ant Group, has released its 1 trillion token-scale AI model, ' Ring-2.6-1T ,' as an open model. Ant Group is a subsidiary of Alibaba, which has multiple AI development teams under its umbrella and has created numerous high-performance AI models such as the Qwen series and Wan series.




Ring-2.6-1T is an inference model with 1 trillion parameters, designed to continuously perform business tasks. The inference strength can be switched between 'high' and 'xhigh,' allowing for flexible use of inference resources by using 'high' for light tasks such as tool calls and task decomposition, and 'xhigh' for scientific research and complex logical analysis.

The graph below shows the benchmark results for 'Ring-2.6-1T', 'Kimi-K2.6 Thinking', 'DeepSeek-V4-Pro Max', 'GPT-5.4 xHigh', 'Gemini-3.1-Pro high', and 'Claude-Opus-4.7 xhigh'. In AIME 26, which measures the mathematical performance of AI, it recorded a score equivalent to DeepSeek-V4-Pro Max.



In ARC-AGI-2, which measures AI intelligence, it significantly outperforms other Chinese-made models, Kimi-K2.6 Thinking and DeepSeek-V4-Pro Max.



In ClawEval, a measure of agent performance, it outperformed GPT-5.4 xHigh and Gemini-3.1-Pro high.



Ring-2.6-1T is distributed as an open model at the following link. It is licensed under the MIT License.

inclusionAI/Ring-2.6-1T · Hugging Face
https://huggingface.co/inclusionAI/Ring-2.6-1T



Related Posts:

in AI, Posted by log1o_hf