GitHub radar
Xing4.0: China Telecom's Agent LLM on Ascend Chips
China Telecom released Xing4.0, a 29B-parameter language model trained entirely on Huawei Ascend NPUs, without NVIDIA hardware.
Xing4.0-29B-A4B is a language model from China Telecom Artificial Intelligence Technology Co., Ltd., published on Hugging Face by the XingChen-AGI organization. It is the successor to their TeleChat series. It uses a Mixture-of-Experts architecture with 29 billion total parameters and 4 billion active per token. The model was trained entirely on Huawei Ascend NPUs using MindSpore. Its native context length is 256K tokens, extendable to 512K. On Terminal-Bench 2.1, which tests agent performance with real terminal tools, it scores 57.5. For comparison, Qwen3.6-35B-A3B scores 51.5 on the same test, and Gemma4-26B-A4B scores 30.0. Its SWE-bench Verified score is 75.0, just below Qwen3.6's 76.0.
Why a vibe-coder should care
What matters here is not only the benchmark numbers, but where this model comes from. China Telecom is a state-owned telecommunications company, not an AI research lab. The fact that they trained a model competitive with Qwen and Gemma of similar size, using only Huawei Ascend hardware, is practically significant. It shows that large language model development is no longer tied to the NVIDIA ecosystem. For agent tasks specifically, the Terminal-Bench score shows a real advantage over similar models.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Open https://huggingface.co/XingChen-AGI/Xing4.0-29B-A4B and tell me how to deploy this model on a server using vLLM — what are the minimum hardware requirements and the basic launch command.
This is a server model. There are no GGUF quantizations available yet. You need a GPU with at least 24 GB of VRAM, or multiple GPUs.
Open on Hugging Face▌ More finds