Under the leadership of He Zhengyu, CTO of Ant Group, the Ling Team has pioneered two open-source Mixture of Experts (MoE) models: Ling-Lite, boasting 16.8 billion parameters, and Ling-Plus, featuring an impressive 290 billion parameters. By integrating AI Infra technology and leveraging domestic AI chips alongside optimized hardware systems, the team has achieved a significant reduction in computational cost by approximately 20%, culminating in a total cost of 5.08 million yuan. This accomplishment places the performance of their models on par with Alibaba's Tongyi Qwen2.5-72B-Instruct and DeepSeek-V2.5-1210-Chat. For the first time, Ant Group has provided a detailed account of its advancements in AI computing power, showcasing its prowess in training large-scale AI models at a lower cost. The related technical paper has been made available on the preprint platform Arxiv.
