Huawei’s AI research team has introduced a new large language model (LLM) using a hybrid technique called Mixture of Grouped Experts (MoGE), designed to enhance training efficiency on its in-house Ascend NPUs.
MoGE builds on the Mixture of Experts (MoE) method used by DeepSeek but addresses key limitations by improving expert load balancing and enabling faster, more efficient parallel execution.
Huawei’s Pangu Pro MoE, featuring 72 billion parameters, reportedly outperforms open-source rivals like GLM-Z1-32B and Qwen3-32B.
The move reflects Huawei’s push to accelerate AI development despite ongoing U.S. sanctions that restrict access to advanced chips and design tools. Nvidia’s CEO recently noted that China’s chip performance is rapidly closing the gap with the U.S.
Source: Mobile World Live
