China Telecom's Xing4.0-29B-A4B: a single-GPU agent model, now available as GGUF

What it's about

China Telecom Artificial Intelligence Technology, or TeleAI for short, is the AI research arm of Chinese telecom giant China Telecom. It has released Xing4.0-29B-A4B, a language model the company describes as a "lightweight agentic large model". "Agentic" means the model shouldn't just answer questions. It should break tasks into steps on its own, call external tools, read long documents and deliver a finished result at the end.

The official press release (via GlobeNewswire, picked up by Business Insider and Yahoo Finance, among others) is dated September 22, 2026. The weights are on Hugging Face and GitHub under the name XingChen-AGI. There's now news for people who run models like this at home: the well-known quantizer bartowski has uploaded ready-made GGUF files. GGUF is the file format used by llama.cpp, LM Studio, Ollama and similar programs, so you can run the model locally without converting it yourself.

How the model is built

Xing4.0-29B-A4B is a mixture-of-experts (MoE) model. Instead of using every part of the network for each word, a kind of router picks just a few specialized subnetworks, called experts. As a result, only about 4 billion of the 29 billion parameters are active per token. That's what the "A4B" in the name refers to. A token is a chunk of a word, the unit language models use to process text. The company says this lets the model answer simple requests almost instantly while still drawing on its full capacity for hard tasks.

bartowski's model card lists 31 billion parameters for the source checkpoint. The gap to the official 29 billion probably comes from extra layers. The model includes MTP (multi-token prediction) layers that predict several upcoming tokens at once. They act as a built-in draft model for speculative decoding. A fast helper guesses several words ahead, and the large model checks them in a single pass. When the guesses are right, text comes out much faster. In llama.cpp, you turn this on with the --spec-type draft-mtp parameter.

The context window, the amount of text the model can keep in view at once, is 256,000 tokens. According to Telecoms Tech News, that's enough to handle entire software repositories, extensive technical documentation and dependencies spread across multiple files. The model card says the model only accepts text. It doesn't understand images.

What the company promises

TeleAI highlights one number above all: 75.0 out of 100 on SWE-bench Verified. In this test, a model has to fix real bug reports from open-source projects. That means digging through code across several files, finding the bug and writing a working patch. The company says this puts the model at the top of its size class. It also claims strong results in tests of function calling, long-context retrieval, code generation and correct JSON output. JSON is a structured data format that programs can process directly. The available sources don't give specific numbers for these tests or a full comparison table against competing models.

China Telecom says the model is already in use inside the company. It runs in the group-wide customer service platform, where it handles complex requests with multi-step reasoning and tool calls, and in assistant systems for home devices. The company says this has "significantly" improved the first-contact resolution rate and handling times. It doesn't give hard figures.

"We believe the future of AI lies not in ever-larger models, but in making powerful intelligence accessible, affordable and deployable everywhere," a China Telecom AI spokesperson said.

The company says the model works with the common open-source frameworks for training, inference and agent development, and it has also been tested on Chinese AI chips. Larger models in the Xing4.0 series are due in the coming months.

The GGUF files in detail

bartowski built the files with llama.cpp release b12174, using an imatrix. This technique uses sample text to measure in advance which parts of the model matter most, then stores those parts more precisely during quantization. The calibration text combines ordinary prose with dialogues full of tool calls and reasoning steps, formatted in the model's chat template.

What's new is a model-specific layout. Instead of fixed default rules, the weights this particular model is most sensitive to get extra precision. This was measured with KL divergence, a measure of how far the quantized model's outputs drift from the original. Lower is better.

Once you account for the difference in file size, the new layout barely helps at Q4 and Q3. At the most aggressive compression, IQ2_XXS, it cuts the drift in half.

A selection of the files:

In every file except Q8_0, the MTP layers are stored as Q4_0, because that's fast and increases the speedup from speculative decoding.

What you need to try it

The company says you need only 15 GB of VRAM. That matches the smallest 3-bit variants, such as IQ3_XXS (14.43 GB) or IQ3_XS (14.92 GB). On a 16 GB card, though, that leaves very little room for context. For the recommended Q4_K_M (20.06 GB), you need a graphics card with 24 GB of VRAM, such as an RTX 3090, 4090 or 5090. A smaller card can still work. Because only 4 billion parameters are active per token, an MoE model often stays reasonably fast when some of the experts sit in system RAM. In that case, your VRAM and RAM combined should be a few GB larger than the file. Running on the CPU alone is also possible, and llama.cpp automatically repacks the weights to suit ARM and x86 processors. Macs with Apple Silicon and at least 32 GB of unified memory are a good fit. bartowski notes that Q4_1 is more efficient per watt there. Plan on 15 to 25 GB of disk space for the file. Important: you need llama.cpp release b12174 or later, or a program built on it.

The bigger picture

Xing4.0-29B-A4B enters a crowded field. MoE models with around 30 billion parameters and only a few active ones are the most popular size for running agents on your own machine. The standout is the claimed SWE-bench Verified score of 75. For a model with 4 billion active parameters, that would be very high. It's unclear how it was measured, for example which agent scaffold was used and how many attempts the model got. There are no independent measurements yet, and the reports from Telecoms Tech News and Business Insider mostly repeat the press release. The sources also don't mention the license, so check it on the original page before any commercial use. bartowski's model card found no details on the prompt format. Even so, if you care about privacy and want to run a coding or research agent entirely on your own machine, the model is worth a try.