China Telecom Open-Sources Xing4.0-29B-A4B Lightweight MoE Agent LLM

China Telecom has open-sourced Xing4.0-29B-A4B, a 29-billion parameter Mixture-of-Experts (MoE) agent model featuring 4 billion active parameters and a native 256K context window. It is the first tens-of-billions parameter model trained entirely on China's domestic hardware and software stack, using Huawei Ascend chips and domestic frameworks. By achieving top-tier agent performance while remaining lightweight enough to run on consumer-grade GPUs, the model demonstrates China's capability to build competitive LLMs independently of Western hardware. Its broad compatibility with mainstream fine-tuning, inference, and agent frameworks makes high-performance AI deployment accessible to individual developers. With 4-bit quantization, the model's VRAM requirement drops to 15GB, enabling long-context processing on consumer graphics cards like the RTX 3090 or RTX 4090. On the SuperCLUE agent evaluation, it scored 93.52—ranking third overall and trailing top Qwen models by less than one point—and supports context expansion up to 512K.