China Telecom AI Open-Sources Xing4.0-29B-A4B: A Fully Domestic Lightweight Code Agent Model

This summer, major tech companies collectively doubled down on AI office work, competing on Agent capabilities, work entry points, and enterprise workflows. But a group of enterprises has not even gotten through the door: data cannot leave their domain, models are best deployed on-premises, and the compute on hand is not exactly abundant—yet long documents must be read and complex business must be handled, and the work that needs doing is no less. China Telecom AI's newly open-sourced Xing4.0-29B-A4B targets precisely this demand.

Its first layer of answer is to lower the deployment threshold. The model uses a MoE architecture with 29 billion total parameters, activating only about 4 billion parameters per inference; under traditional FP16 precision, a 29B model occupies about 60GB of single-card memory, often requiring multiple cards or expensive A-series cards, but after 4-bit quantization this usage can be compressed to about 15GB, a reduction of about 75%, allowing consumer-grade GPUs such as the RTX 3090 and 4090 to run it.

The second layer of answer is to make the entire technical chain fully domestically produced: from domestic chips and domestic training frameworks to model training and inference deployment, systematic adaptation has been done. The model is trained entirely on Huawei Ascend 910C compute, uses the domestic MindSpore/MindFormers training framework and development kit, and has completed Ascend adaptation verification on the inference deployment side.

Tasks handed over by enterprises are rarely a single question-and-answer. After reading a document, one may still need to extract information, call tools, verify results, and finally organize them into deliverable materials; this work is done every day, and it must be both stable and cost-effective. This is exactly why Xing4.0-29B-A4B focuses on strengthening Coding and Agent capabilities.

In Coding, it pushes capabilities from writing snippets to doing engineering: it natively supports 256K context and can be extended to 512K, allowing a single task to hold longer code, documents, logs, and historical records. The article gives an example: turning contract ledgers scattered across Excel into a management dashboard requires not just charting functions, but also understanding the tabular business data and organizing contract overview, amount distribution, risk identification, and performance warnings onto the same page, and the model can do this locally within the enterprise, with data never leaving the domain throughout.

In Agent, the model must break down tasks itself, arrange execution order, call tools, and then decide the next step based on intermediate results until an acceptable result is delivered. To plug into existing workflows, it has completed targeted adaptation for mainstream Agent and Harness frameworks such as OpenCode, Claude Code, OpenClaw, and Hermes, while also being compatible with commonly used toolchains such as LLaMA-Factory, MindFormers, SGLang, vLLM, and KTransformers.