China Telecom AI releases Xing4.0-29B agentic model for single-GPU deploy

China Telecom Artificial Intelligence Technology Co., Ltd. (China Telecom AI, also called TeleAI) released Xing4.0-29B-A4B, a lightweight agentic model with 29 billion parameters in total and 4 billion turned on, and says it runs on one consumer graphics card at 15 GB of memory with a 256,000-token context window.

SOFTWARE desk — an agent that fits on one consumer graphics card changes who can run tool-calling on their own machine, without a server cluster. Tool-calling means the model can use outside software, not only reply in chat.

What kind of model the release says it is. Xing4.0-29B-A4B uses a Mixture-of-Experts design, shortened to MoE. That is a model made of specialist blocks, with only some of them turned on for each piece of input. The release says that is why 4 billion of the 29 billion parameters are activated per token. A token is a small chunk of text, roughly a word piece. The release says the model is built for execution. It can plan a multi-step path, call an outside tool, take a long input, and hand back a finished result. A tool, here, is software the model is allowed to use, such as a code runner. The context window is 256K tokens. A context window is how much text the model can hold at once. 256K is about 256,000 tokens, enough for a long set of documents. The release says that window is for engineering work that has to reason across many documents. File the design, the per-token split, the tool use, and the 256K window as the release’s.

The speed line stays the company’s. The release says turning on only 4 billion parameters per token lets the model answer a simple question quickly and still use its full size for a hard multi-step task. “Near-instantaneous” and “optimal balance” are the release’s words. This desk did not time a reply.

The single-card claim. The release says low-bit quantization and memory optimization bring the model to only 15 GB of GPU memory, so it can run locally on one consumer-grade graphics card. Quantization, here, means storing the model’s numbers with fewer bits so they take less memory. 15 GB is the memory of a higher-end consumer card, not a rack of data-center chips. The release says that removes the need to buy a multi-GPU server cluster to build an agent app. It also says a person can run code work, data analysis, and document work on their own machine, with no data leaving the device. That privacy line is the company’s. This desk did not load the weights and did not measure the memory.

The score stays labeled. On SWE-bench Verified, the release says Xing4.0-29B-A4B scored 75.0 out of 100 and ranks among the top models in its parameter class. SWE-bench is a test of whether a model can fix real software issues. Verified is the checked slice of that test. 75.0 out of 100 is the company’s figure. “Among the top” is the company’s ranking, not a board this desk compiled. The release also says the model is strong at multi-step agent tasks, function-calling accuracy, long-context retrieval, JSON format compliance, and code generation. Function calling means the model picks a tool and fills in the fields. JSON is a structured text format programs use to pass data. The release does not print those other scores. This desk did not rerun SWE-bench.

Where China Telecom says it already runs. The release says the model is in the company’s group-level customer-service platform, handling complex questions with multi-step reasoning and tool calling. It says that work has significantly improved first-contact resolution and average handling efficiency. It does not print a rate or a minute count. Significantly is the company’s word. The release also says the model is deployed in “mid-screen” interactive service scenarios, a real-time assistant for customers in home-service settings. File both deployments, and the improvement line with no number on it, as the release’s. This desk did not watch a call.

Where the weights are. The release says China Telecom AI has put Xing4.0-29B-A4B on GitHub and Hugging Face under the name XingChen-AGI. The Hugging Face link printed on the page is https://huggingface.co/XingChen-AGI. GitHub is named as XingChen-AGI, without a full repository address on the page. Weights are the downloadable model files. The release says the model works with mainstream open-source training, inference, and agent frameworks, and that it has been checked on a range of domestic and mainstream AI chips. It does not name the chips. It says larger Xing4.0 models are planned in the coming months. This desk did not clone a repository and did not name a chip the page left blank.

The spokesperson line. A spokesperson for China Telecom AI says the future of AI is not ever-larger models. It is intelligence that is accessible, affordable, and deployable anywhere. The spokesperson says Xing4.0-29B-A4B shows a compact model can deliver agentic capability on a developer’s desktop or in an enterprise system. That is the company’s framing. It is not a measurement.

Do not fill blanks the page left open. The release says “recently officially released” on a 22 Sep 2026 stamp. It does not print an earlier date. Do not turn 15 GB into a card this desk tested. Do not turn 75.0 into an independent ranking. Do not turn “significantly improving” into a percentage. Do not name a chip, a license, or a GitHub address the page did not print.

Plain English for the rest of the card: China Telecom AI / TeleAI = China Telecom Artificial Intelligence Technology Co., Ltd. Xing4.0-29B-A4B = the model in the release. 29B = 29 billion parameters, the whole model. 4B active = about 4 billion parameters turned on for each token. MoE = mixture of experts, specialist blocks with only some on at once. token = a small chunk of text. 256K context = room for about 256,000 tokens at once. GPU = the graphics chip. 15 GB = the memory the company says the optimized model needs. quantization = storing the numbers with fewer bits. SWE-bench Verified = a checked test of software-issue fixes. 75.0 = the company’s score out of 100. tool calling = the model uses outside software. JSON = structured text for programs. weights = the downloadable model files. XingChen-AGI = the name on GitHub and Hugging Face. This filing is the 22 Sep release.

PRIMARY here: China Telecom AI’s 22 Sep 2026 GlobeNewswire release, “China Telecom AI Officially Releases Xing4.0-29B Agentic Large Model for Single-GPU Deployment” — Tier A PRIMARY, the company’s own record. The 01:55 ET stamp, the Beijing dateline, the 29 billion / 4 billion split, the Mixture-of-Experts line, the 256K context, the 15 GB single-card claim, the SWE-bench Verified 75.0, the customer-service and mid-screen deployments, the XingChen-AGI GitHub and Hugging Face lines, and the spokesperson quote are the release’s. NOT claimed: that this desk loaded the weights, measured 15 GB, reran SWE-bench, watched a customer call, named a chip, or confirmed a GitHub address the page did not print, a stock tip, or investment advice. Distinct from the already-filed alibaba-zhenwu-v900, deepseek-unsc-ai-brief, and xiaomi-mimo-v2-6.