Ant Group has developed AI models using Chinese-made chips from Alibaba and Huawei, alongside AMD alternatives, cutting costs by about 20%.

While still using some Nvidia chips, the company aims to reduce reliance on high-performance GPUs like Nvidia’s H800, which face US export restrictions.

Its AI models, based on the Mixture of Experts (MoE) technique, reportedly achieve results similar to Nvidia hardware, though independent verification is lacking.

Ant’s research highlights its Ling-Plus and Ling-Lite models, which outperformed Meta’s Llama in English benchmarks and DeepSeek in Chinese tests. Open source, Ling-Plus has 290 billion parameters, while Ling-Lite has 16.8 billion.

The company estimates training one trillion tokens with its optimized approach costs 5.1 million yuan (US$880,000), significantly less than traditional high-performance hardware.

Despite progress, its research notes challenges in model training, including stability issues from hardware or structural adjustments.

Recent Ant Group developments