Ant Group Open-Sources Trillion-Parameter AI Models Ling-2.5-1T and Ring-2.5-1T

Ant Group Goes Big on Open Source AI

Ant Group, the Chinese fintech powerhouse behind Alipay, has released two new trillion-parameter AI models under open licences, marking a significant escalation in the global AI arms race. The models, Ling-2.5-1T and Ring-2.5-1T, are now freely available on Hugging Face and ModelScope.

Efficiency Meets Scale

Ling-2.5-1T serves as Ant Group's flagship non-reasoning model, featuring one trillion total parameters with sixty-three billion active parameters and support for context lengths of up to one million tokens. What sets it apart is its remarkable token efficiency. On the AIME 2026 maths benchmark, the model matched the performance of leading thinking models that typically require between fifteen thousand and twenty-three thousand tokens, while using only around five thousand eight hundred and ninety tokens. This efficiency comes from a hybrid linear attention architecture trained on twenty-nine trillion tokens of pre-training data.

A Thinking Model That Does Gold-Medal Maths

Ring-2.5-1T is described as the first open-source trillion-parameter thinking model built on hybrid linear architecture. It scored thirty-five out of forty-two on the International Mathematical Olympiad 2025 test, matching the gold medal cutoff, and achieved one hundred and five out of one hundred and twenty-six on the Chinese Mathematical Olympiad 2025, surpassing China's national team cutoff. The architecture also reduces memory access by more than ten times and increases generation throughput by more than three times compared to its predecessor.

Building an Open AI Empire

The releases extend Ant Group's BaiLing model family, which now includes three main series: Ling non-thinking models, Ring thinking models, and the multimodal Ming series. Ant Group's AI subsidiary, InclusionAI, has released eighteen models and reached trillion-parameter scale over the past year as part of its push toward artificial general intelligence.