GitHub radar
China Telecom’s Agent LLM powered by Ascend chips.
China Telecom has developed Xing4.0 (a 29B param language model) with Huawei Ascend NPUs only (no NVIDIA GPUs).
Xing4.0-29B-A4B is a language model developed by China Telecom Artificial Intelligence Technology Co., Ltd., made available on Hugging Face by the XingChen-AGI organization. It is a follow up to their TeleChat family of models. It employs a Mixture-of-Experts structure, with a total of 29 billion parameters and 4 billion effective params per token. It was trained on Huawei Ascend NPUs using MindSpore. It has a default context window of 256K tokens with an option to extend it up to 512K. In the Terminal-Bench 2.1 benchmark, which evaluates agents on a suite of terminal utilities, this model achieves a score of 57.5. For reference, Qwen3.6-35B-A3B achieves 51.5 on the same benchmark, while Gemma4-26B-A4B achieves 30.0. This model’s SWE-bench Verified score is 75.0, compared to Qwen3.6’s 76.0.
Why a vibe-coder should care
The important thing about this model isn’t so much its performance in benchmarks but its origin. China Telecom is not an AI research institution — it’s a state-run telecom provider. That they were able to train a model that performs on par with Qwen and Gemma (of comparable size), using only Huawei Ascend hardware, has major implications. It demonstrates that you don’t need NVIDIA GPUs to build large language models. Specifically, in the realm of agents, the Terminal-Bench score indicates a clear improvement over other 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