Compact footprint, full-scale capability—China’s lightweight agentic large model runs locally on consumer-grade hardware while also powering enterprise production systems
BEIJING, Sept. 22, 2026 (GLOBE NEWSWIRE) — China Telecom Artificial Intelligence Technology Co., Ltd. (China Telecom AI) recently officially released Xing4.0-29B-A4B, its next-generation lightweight agentic large model. With 29 billion total parameters and just 4 billion activated parameters, the model provides enterprise-grade agentic capabilities at an extremely low compute threshold: it can be deployed in large-scale enterprise production environments or run locally on a personal consumer-grade graphics card, offering developers and enterprises a highly cost-effective AI agent solution.
Built on a next-generation Mixture-of-Experts (MoE) architecture, Xing4.0-29B-A4B is purpose-built for the agentic era. Unlike conventional interactive models that focus on question-and-answer exchanges, Xing4.0-29B-A4B places greater emphasis on execution. It can independently plan multi-step task paths, call external tools, process long-context inputs, and deliver completed results. With an ultra-long 256K-token context window, the model is well equipped for complex engineering tasks that require sustained reasoning across a large number of documents.
Speed Meets Efficiency
Xing4.0-29B-A4B strikes an optimal balance between inference speed and task complexity. By activating only 4 billion of its 29 billion parameters per token, the model delivers near-instantaneous responses to simple queries while still being able to draw on its full capacity for highly demanding multi-step tasks, combining the deep reasoning power of a large model without sacrificing the responsiveness of a lightweight one.
Single-GPU Deployment: From Data Center to Desktop
Using low-bit quantization and memory optimization techniques, Xing4.0-29B-A4B requires only 15 GB of GPU memory, which means it can run locally on a single consumer-grade graphics card. This dramatically lowers the infrastructure barrier for developers and enterprises building AI agent applications, eliminating the need to invest in expensive multi-GPU server clusters.
For individual developers and technology enthusiasts, Xing4.0-29B-A4B delivers a fully local AI agent experience with complete privacy control. Users can run sophisticated agent workflows — including code development, data analysis, and document processing — entirely on their personal devices, with no data leaving their device. This is particularly valuable for organizations with strict data privacy requirements, as well as for developers who need to validate prototypes before scaling to full deployment.
Compact Size, Complete Capabilities
Despite its small size, Xing4.0-29B-A4B offers complete agentic capabilities. The model performs strongly across a range of core benchmarks, covering multi-step agent task completion, function-calling accuracy, long-context retrieval, JSON format compliance, and code generation. On the SWE-bench Verified benchmark, Xing4.0-29B-A4B scored 75.0 out of 100, ranking among the top models in its parameter class.
Unlike other lightweight models, Xing4.0-29B-A4B can handle complex, real-world engineering scenarios. It can autonomously break high-level user goals down into actionable steps, read and analyze multi-file project repositories, execute code, and generate structured deliverables — capabilities that until now were available only in large models with far more parameters.
Proven in Production
Xing4.0-29B-A4B has been deployed and validated in real-world enterprise environments. China Telecom has integrated the model into its group-level customer service platform, where it handles complex customer inquiries through multi-step reasoning and tool-calling workflows, significantly improving first-contact resolution rates and average handling efficiency. The model is also deployed in “mid-screen” interactive service scenarios, providing real-time, context-aware intelligent assistant services to customers in home service environments.
These production deployments demonstrate that Xing4.0-29B-A4B is not only a research benchmark but also a battle-tested, production-ready model that delivers measurable business value for customers.
Open Ecosystem
Following its open-ecosystem strategy, China Telecom AI has released Xing4.0-29B-A4B on major open-source communities such as GitHub and Hugging Face. The model is compatible with mainstream open-source training, inference, and agent development frameworks, and has been validated across a range of domestic and mainstream AI chip platforms.
“We believe the future of AI lies not in ever-larger models, but in making powerful intelligence accessible to everyone, affordable, and deployable anywhere,” said a spokesperson for China Telecom AI. “Xing4.0-29B-A4B proves that a compact model can also deliver agentic capabilities on par with far larger models — whether on a developer’s desktop or in an enterprise production environment.”
Availability
Xing4.0-29B-A4B is now officially available, and developers can access the model through the following platforms:
China Telecom AI said it will continue to advance the Xing4.0 model series and plans to release larger-parameter models in the coming months, further pushing the boundaries of what compact, efficient AI architectures can achieve in real-world applications.
About China Telecom AI (TeleAI)
China Telecom AI (TeleAI) serves as the core artificial intelligence research and development division of China Telecom, dedicated to advancing foundational AI and practical industrial applications. Leveraging a robust cloud-network infrastructure, TeleAI develops scalable solutions like the Xingchen foundation model series. These advanced models currently power over 500 enterprise workflows across more than 20 industries worldwide. Through continuous innovation, China Telecom AI is committed to driving digital transformation and establishing AI as a primary engine for modern productivity.
Website: https://www.teleai.com.cn/
Contact
Xing Chen
DataAiTech-service@chinatelecom.cn
