Ant Group Unveils Two Groundbreaking MoE Large Models, Drastically Cutting Training Costs
2025-03-24 / Read about 0 minute
Author:小编   

The Ling team at Ant Group has released a groundbreaking paper, unveiling two novel large language models: the Bailing Lightweight Edition and the Enhanced Edition, boasting parameter scales of 16.8 billion and 290 billion, respectively. The team has implemented innovative training methodologies, including dynamic parameter allocation and mixed precision scheduling, enabling efficient training even on lower-performance hardware, thereby drastically reducing costs. Experimental results have demonstrated that post-optimization, the training expenses utilizing lower-specification hardware have been lowered by approximately 20%, while maintaining performance on par with industry-leading models. This represents a highly cost-effective solution for domestic large models and mitigates the reliance on high-end chips.