2025 Wuzhen Internet Summit: Ant Group has deployed a domestic computing power group with a scale of ten thousand cards

[Global Network Technology Report Reporter Burchan] On November 8, at the 2025 World Internet Conference and Wuzhen Summit Frontier Artificial Intelligence Model Forum, Luo Ji, presidentof Ant Group'sPlatform Technology Business Group, revealed that in terms of computing power, ants have deployed a domestic computing power cluster of 10,000 cards. It adapts to self-developed and various mainstream open source models. The stability of training tasks exceeds 98%. The training and reasoning performance is comparable to that of international computing power clusters, and is fully applied to large model training and reasoning services in the field of security risk control.
Luo Ji said that in recent years, General Artificial Intelligence (AGI) has developed rapidly driven by the "Scaling Law", and some capabilities have reached and exceeded human benchmarks. The training data volume of mainstream flagship language large models has exceeded 20T, and the parameter scale has entered the "trillion" era. At the same time, the shortage of computing power resources and rising costs have become key factors that restrict AGI from continuing to break through under the "Law of Scale". factor.
In this context, Ant Group has explored a series of technological innovations to achieve breakthroughs in intelligent experience, including the optimization of the "three major efficiencies" of model parameter efficiency, data application efficiency, and computing efficiency, and leveraging the "two key factors" of intelligent experience."-the response speed of the model and the scale of the model's knowledge, hoping to promote the inclusive application of AGI.
Luo Ji introduced the effectiveness of these technological innovations using the ant lark model as an example. In the past, Ant Bailing has built a large model system with full size, full modality, and comprehensive open source covering language, reasoning, and multimodal. Among them, the world's first open source trillion-dollar thinking model Ring-1T released in October demonstrated leading logical reasoning and coding capabilities. Its mathematical ability reached the silver medal level of the International Mathematical Olympics. It also performed in medical health and creative writing. Strong performance, behind it, innovative training methods have stabilized the length of both training and reasoning and nearly doubled the training efficiency.
It is worth mentioning that ants have also made important breakthroughs in controlling the number of Tokens generated by reasoning. Ling-1T adopts an evolutionary thinking chain to significantly reduce token resource consumption while ensuring correct thinking rates, achieving task effectiveness and calculation."Pareto optimality" between costs.
Luo Ji emphasized: "The ideal future is not that AI replaces people, but that AI helps humans amplify their capabilities and achieve a high degree of human-machine integration and collaboration. In the era of coexistence between man and AI, we must not only continuously improve the level of intelligence, but also ensure that AI technology follows ethics, safety and credibility."
Sina statement: All meeting minutes are compiled in shorthand on the spot and have not been reviewed by the speaker. Sina posted this article for the purpose of conveying more information and does not mean that they agree with their views or confirm their descriptions.