China Telecom Opens Xing4.0 Agent Model Weights
China Telecom released Xing4.0-29B-A4B, an Apache-licensed sparse model trained on Ascend hardware and built for long-context agent workflows.
China Telecom’s AI unit has released Xing4.0-29B-A4B, a 29-billion-parameter mixture-of-experts model with approximately 4 billion parameters active per token. The model card is available on Hugging Face under the Apache-2.0 license.
Xing4.0 supports a native 256,000-token context window, with extension to 512,000 tokens, and is positioned for coding, planning, tool use, and other agent workflows. Its architecture combines multi-head latent attention, hyper-connections, multi-token prediction, and 64 routed experts, of which four are selected for each token alongside a shared expert.
The release is notable for its hardware and software stack. China Telecom says the model was trained entirely on Huawei Ascend 910C NPUs with the MindSpore framework. The model card describes optimizations for expert communication, recomputation, graph fusion, and Ascend-specific kernels, reporting roughly 96% higher training throughput than an out-of-the-box setup. Those figures are vendor-reported and should not be read as a general comparison with NVIDIA-based training.
The weights are compatible with Transformers, vLLM, SGLang, KTransformers, LLaMA-Factory, and MindFormers, although support for the architecture is still uneven across inference runtimes. The model card also lists integrations with several coding and agent frameworks.
Independent evaluations suggest a mixed profile. Xing4.0 performs competitively on some agent and terminal benchmarks, while trailing stronger comparison models on several reasoning and software-engineering tests. That pattern fits the model’s stated emphasis: practical tool execution and long context rather than universal frontier leadership.
Why it matters: this is both a usable open model and a public demonstration that a major Chinese telecom operator can expose an end-to-end domestic training stack. Its long-term importance will depend on runtime maturity, reproducibility, and whether developers can turn the advertised agent capabilities into reliable private deployments.
Uncle Cat take
Xing4.0’s strongest signal is not its 29B headline but the Ascend-native path; deployment friction will decide whether that stack travels beyond China Telecom.