China Telecom Releases New Large Language Model Xing4.0-29B-A4B

China Telecom Artificial Intelligence Technology Co., Ltd. officially released its new generation large language model, Xing4.0-29B-A4B, on September 17. The model focuses on capabilities such as complex task understanding, task planning, tool invocation, and autonomous execution. It supports long context processing and is applicable to scenarios including code development, data analysis, office automation, and life services.

The model has a total of 29 billion parameters with 4 billion active parameters. It is reported to be the first ten-billion-parameter model trained in China using Huawei's Ascend Atlas 900 A3 SuperPoD liquid-cooled supernode and the MindSpore framework. The model has been deeply optimized for complex engineering tasks.

Huawei stated that the model employs a new architecture to enhance long-range task processing and multi-step execution. It can autonomously plan task paths, invoke tools, and complete complex tasks based on user-defined objectives.

The training was conducted on Huawei's Ascend supercomputing cluster, achieving a 96% improvement in training performance. The process involved extensive data collection, architecture optimization for long contexts, and multi-expert reinforcement learning.