China Telecom Releases Xing4.0-29B-A4B AI Model Trained on Huawei Ascend

China Telecom AI Technology released its new Xing4.0-29B-A4B large model on September 17. The model focuses on complex task understanding, planning, tool use, and self-execution, with support for long context processing. It can be used for coding, data analysis, office automation, and daily services. The model has 29B total parameters with only 4B active.

Huawei announced this is China's first 10-billion-parameter model trained on Ascend Atlas 900 A3 SuperPoD liquid-cooled super nodes and MindSpore framework. It uses MLA attention compression, multi-token prediction, and mHC architecture, with training sequences scaling from 4K to 256K.

Engineering optimizations boosted training performance by 96%, MoE throughput by 32%, and custom Ascend C operator efficiency by 30%. The model uses Slime reinforcement learning across Agent, Coding, and Reasoning tasks. Ascend now supports native training for over 40 leading industry models.