China telecom launches open-source lightweight AI model Xing4.0-29B-A4B

China Telecom has released the Xing4.0-29B-A4B, its first fully open-source, lightweight AI model built on domestic infrastructure. The model, with 29 billion parameters and 4 billion active parameters, supports up to 512K context length, making it the first domestic large model trained and optimized for complex engineering tasks using indigenous computing power and frameworks. According to SuperCLUE evaluations, the model scored 93.52, ranking third, with a gap of less than one point from top-tier Qwen models. The model, optimized with 4-bit quantization, requires only 15GB of memory, enabling it to run on consumer-grade GPUs like the RTX 3090 and RTX 4090. It is fully compatible with major open-source ecosystems, including LLaMA-Factory, MindFormers, SGLang, vLLM, and KTransformers. The model is available on HuggingFace, GitHub, and other platforms.