On September 19, China Telecom's AI subsidiary released Xingchen (Xing) Xing4.0-29B-A4B, an MoE model with 29B total parameters and only 4B activated, described as the first hundred-billion-parameter model in China trained entirely on domestic compute and a domestic framework and optimized for complex engineering tasks.
The model natively supports a 256K context window, extensible to 512K, and after 4-bit quantization runs on consumer GPUs such as RTX 3090/4090 (24GB) for long-context tasks. On the SuperCLUE agent capability benchmark it scored 93.52, third overall and within 1 point of two leading Qwen models. It has been open-sourced on GitHub, Hugging Face, Gitee, ModelScope and MoLing.
The release highlights a trend toward lightweight, domestically trainable models that bring agent capabilities to ordinary hardware, lowering the barrier for enterprise deployment of Chinese foundational models.
Source: People's Posts and Telecommunications News / China Telecom
