Ant Group's Ling team unveiled two MoE models: Ling-Lite (16.8B params) & Ling-Plus (290B params), delivering industry-leading performance. Trained efficiently on China's domestic GPUs, rivaling NVIDIA-based models.

It's noted that the Ant Group team utilized lower-spec hardware systems during the model pre-training phase, reducing computation costs by approximately 20%, amounting to 5.08 million yuan, lower than DeepSeek's 5.576 million yuan.