Ant Group Unveils Ring-1T, a Trillion-Parameter AI Model

Ant Group has unveiled Ring-1T-preview, an open-source artificial intelligence model with more than one trillion parameters, released on Hugging Face.

It is the first open-source model of its scale specifically designed for natural language reasoning, marking a milestone in the global AI race.

Built on the Ling-2.0 Mixture-of-Experts (MoE) architecture and pretrained on more than 20 trillion tokens, Ring-1T-preview is engineered to push the frontiers of reasoning and problem-solving tasks. Early benchmarks show the model’s strong performance across multiple domains. On the AIME 2025 reasoning test, Ring-1T-preview scored 92.6, surpassing other open-source models and even outperforming Gemini 2.5 Pro, while closely approaching GPT-5’s 94.6 benchmark.

Key Highlights

Ant Group launches Ring-1T-preview, a trillion-parameter AI model
Model scores 92.6 on AIME 2025, outperforming Gemini 2.5 Pro
Open-sourced on Hugging Face, built on Ling-2.0 MoE architecture

The model also topped leaderboards on CodeForces, LiveCodeBench, and ARC-AGI-v1, reflecting robust capabilities in programming, logic, and general intelligence benchmarks.

By making the trillion-parameter model open-source, Ant Group signals its commitment to driving AI accessibility and innovation. The release also underscores the growing strategic emphasis by Chinese tech giants on competing in the AI infrastructure space, while enabling global developers and enterprises to build advanced applications on top of state-of-the-art reasoning models.

Also Read: Ant Group, Peking University Partner on AI Healthcare Lab

With Ring-1T-preview, Ant Group positions itself as a serious contender in next-generation AI development, directly engaging with global leaders in the race toward more capable and reliable artificial general intelligence (AGI).