Ant Group Launches World’s First Open-Source Trillion-Parameter Reasoning Model

Breaking NewsSep. 30, 2025

Summarized by NextFin AI

Ant Group has launched its first trillion-parameter AI model, Ring-1T-preview, on Hugging Face, marking a significant milestone as the world’s first open-source reasoning model of this scale.
The model achieved a benchmark score of 92.6 on AIME 25, outperforming all known open-source systems and Google’s Gemini 2.5 Pro, and is close to GPT-5’s score of 94.6.
On CodeForces, Ring-1T-preview scored 94.69, surpassing GPT-5, and led in various benchmarks including LiveCodeBench and ARC-AGI-v1.
Built on the Ling-2.0 MoE architecture and trained on 20T tokens, the model utilizes Ant’s icepop method and its proprietary RLVR system ASystem, with components like AReaL already open-sourced.

AsianFin -- Ant Group has unveiled its first trillion-parameter AI model, Ring-1T-preview, on Hugging Face, marking the world’s first open-source reasoning model at this scale.

The model posted standout results in benchmarks, scoring 92.6 on AIME 25—beating all known open-source systems and Google’s Gemini 2.5 Pro, and approaching GPT-5’s 94.6 (without tools). On CodeForces, it scored 94.69, surpassing GPT-5, and topped leaderboards such as LiveCodeBench and ARC-AGI-v1.

Built on the Ling-2.0 MoE architecture and trained on 20T tokens, Ring-1T leverages Ant’s icepop method and its self-developed RLVR system ASystem, with parts like AReaL already open-sourced. The model remains under training.

Explore more exclusive insights at nextfin.ai.

Insights

What is a trillion-parameter AI model and how does it differ from smaller models?

How did Ant Group develop the Ring-1T-preview model?

What are the key features of the Ling-2.0 MoE architecture?

What benchmarks does the Ring-1T-preview model outperform other models in?

What is the significance of open-sourcing AI models in the tech industry?

How does the performance of Ring-1T-preview compare to Google's Gemini 2.5 Pro?

What feedback have users given about the Ring-1T-preview model so far?

What are the implications of Ant Group's new model for the future of AI development?

What challenges did Ant Group face in creating the world's first trillion-parameter model?

How does the icepop method contribute to the performance of Ring-1T?

What are the potential ethical concerns surrounding the use of trillion-parameter models?

How does the self-developed RLVR system ASystem enhance the capabilities of Ring-1T?

What industry trends are emerging in the development of large AI models?

What are the potential applications for the Ring-1T-preview model in real-world scenarios?

How does the open-source nature of Ring-1T-preview affect competition among AI companies?

What historical precedents exist for the development of large-scale AI models?

How might the introduction of models like Ring-1T influence future AI research?

What are some limitations associated with trillion-parameter AI models?

In what ways could the Ring-1T-preview model revolutionize AI reasoning tasks?

NextFin.Al

No Noise, only Signal.

Open App