China Telecom AI Releases 29B Agent Model for Single-GPU Deployment

China Telecom AI has released Xing4.0-29B-A4B, a downloadable model aimed at bringing multi-step AI agent work to a single consumer graphics card. The practical proposition is not merely a smaller model: China Telecom AI says developers can run a system designed to plan tasks, call tools and work across long documents without a multi-GPU server setup.

The release is available through the XingChen-AGI GitHub organization and Hugging Face. It uses a mixture-of-experts design: the model has 29 billion parameters in total, but activates 4 billion for each token it processes. Its total size is therefore different from the share active during a response.

29BTotal parameters

Xing4.0-29B-A4B has 29 billion total parameters.

4BParameters activated per token

Its mixture-of-experts architecture activates 4 billion parameters per token.

15 GBStated GPU-memory requirement

China Telecom AI says quantization and memory optimization let the model run on one consumer-grade graphics card using about 15 GB of GPU memory.

A model aimed at local agent workflows

China Telecom AI says low-bit quantization and memory optimization bring the stated memory requirement to roughly 15 GB. The company presents that setup as an alternative to expensive multi-GPU server clusters for development or deployment.

The model has a stated 256,000-token context window, the amount of text it can take into a request at once. China Telecom AI says it is intended for complex engineering work spanning many documents, rather than short question-and-answer exchanges.

The tasks China Telecom AI says it targets

• Planning a multi-step path from a high-level goal to completed work.
• Calling external tools as part of that workflow.
• Handling code development, data analysis and document processing locally.

Production use is part of the company’s case

China Telecom says it has integrated Xing4.0-29B-A4B into its group-level customer-service platform, where the model handles complex inquiries with multi-step reasoning and tool calls. The company also reports deployments in interactive home-service settings. These are company-reported production examples, rather than independently measured evidence of reliability or efficiency.

The company also reports a 75.0 score on SWE-bench Verified and says that places the model among the top systems in its parameter class. It says the model works with mainstream open-source training, inference and agent-development frameworks. Both claims are part of the release’s own performance and compatibility case.

Local control, with a test still ahead

For teams with sensitive documents or data, the appeal is that an agent workflow could run on their own device. China Telecom AI says local operation means data need not leave that device. The remaining question is whether its stated memory efficiency and agent performance will hold across the workflows developers choose to run.