China Telecom Releases Xing4.0-29B-A4B, a Fully Domestic Large Model

China Telecom has released Xing4.0-29B-A4B, a new large model with 29B total parameters and only 4B active parameters. It natively supports 256K context and can extend to 512K. It is China's first hundred-billion-parameter model trained entirely on domestic chips and frameworks.

The company says the entire stack is made in China, from training chips to inference deployment, using Ascend chips, domestic frameworks, self-developed architecture, and open-source ecosystem. On the SuperCLUE benchmark, it scored 93.52 in agent capabilities, ranking third and less than 1 point behind two top Qwen models.

After 4-bit quantization, memory usage drops to 15GB, meaning 24GB consumer GPUs like the RTX 3090 and RTX 4090 can run long-context tasks locally. The model stays open-source and works with LLaMA-Factory, MindFormers, SGLang, vLLM, and other tools. It also has tuned support for agent frameworks like OpenCode, Claude Code, and OpenClaw.

The model is now available on HuggingFace, GitHub, ModelScope, Gitee, and Modelers.