With 29 billion parameters, only 4 billion activated per token, and training throughput improved by roughly 96%, China Telecom Artificial Intelligence Technology’s open-source release of Xing4.0 matters for more than leaderboard rankings. It shows that China’s AI infrastructure is beginning to form a complete stack spanning frameworks, computing power, and inference tools.
What this is
Xing4.0-29B-A4B is the new large language model from China Telecom Artificial Intelligence Technology. It uses a sparse mixture-of-experts architecture: the model has 29 billion total parameters, but activates only about 4 billion for each token, reducing inference computation. It natively supports a 256,000-token context window and can be extended to 512,000 tokens.
The model was trained on a Huawei Ascend 910C cluster using the MindSpore framework. It is optimized for multi-step planning, tool calling, and long-running task execution. The model also supports tools such as LLaMA-Factory, SGLang, and vLLM, making it suitable for enterprise use cases including intent classification, table understanding, contract review, and knowledge Q&A.
Industry view
This release sends a clear signal: competition among China’s domestic models is shifting from parameter scale toward hardware-software co-optimization. When chips, frameworks, kernels, and deployment tools can be optimized together, a model’s actual cost of use may matter more than its theoretical parameter count.
The risks are just as clear. The roughly 96% improvement in training efficiency was measured against an out-of-the-box solution; it does not mean every enterprise deployment will see the same gains. The model’s accuracy, stability, licensing boundaries, and maintenance costs in real business operations still require independent evaluation. Ultra-long context also does not guarantee that every long task will be completed reliably. When tool calling involves high-risk operations such as payments or deleting and modifying data, enterprises still need to enforce permissions and retain human review.
Impact on regular people
For enterprise IT: The model’s deep integration with the Ascend ecosystem gives enterprises another option for private deployments where data cannot leave the organization. However, buyers should compare the full cost of completing tasks rather than looking only at parameter counts.
For workplace productivity: Tasks such as contract review, spreadsheet analysis, and knowledge retrieval may gradually be integrated into localized agent tools. In the near term, the realistic outcome is faster workflows, not wholesale job replacement.
For the consumer market: If the open ecosystem continues to expand, agent services may become better aligned with domestic chips and local deployment needs. The experience of individual products, however, will still depend on how application teams integrate tools, design security, and deliver service quality.