China Telecom open-sources Xing4.0-29B-A4B, a 29B MoE coding agent that runs on one RTX 3090

On September 20, China Telecom AI open-sourced Xing4.0-29B-A4B, a lightweight coding-agent model it describes as fully domestic across its entire technical stack. The launch targets a specific problem: enterprises whose data cannot leave their networks and whose compute budgets are tight, but which still need to read long documents and run complex workflows locally.

The parameter layout is the headline. Xing4.0-29B-A4B uses a mixture-of-experts architecture with 29 billion total parameters and only about 4 billion active per inference pass. After 4-bit quantization, memory use drops from roughly 60GB in FP16 to about 15GB, a reduction of about 75%, which puts it within reach of consumer cards such as an RTX 3090 or 4090.

Fully domestic refers to the training path. The model was trained on Huawei Ascend 910C clusters using the domestic MindSpore and MindFormers frameworks, and its inference deployment has also been validated on Ascend hardware, so neither training nor serving depends on foreign accelerators.

The model natively supports 256K context and can be extended to 512K. To make long context practical, it uses multi-head latent attention to compress the KV cache, multi-token prediction to speed up generation, and mHC manifold-constrained hyper-connections to curb numerical instability during training.

Capability work centres on coding and agents. China Telecom frames the upgrade as moving from writing code fragments to doing engineering work, such as turning contract ledgers scattered across spreadsheets into an on-premise management dashboard, and strengthening long-horizon tasks where the model decomposes steps, sequences them, calls tools and adjusts based on intermediate results.

Fitting into existing environments is another pillar. The model ships with targeted adaptations for mainstream agent and harness frameworks including OpenCode, Claude Code, OpenClaw and Hermes, and is compatible with toolchains such as LLaMA-Factory, MindFormers, SGLang, vLLM and KTransformers, lowering the cost of swapping it into a live development setup.

Deployment numbers were part of the pitch. In a customer-service deployment the model runs fully privately, keeping call audio, identity data and business records inside the enterprise, with the vendor claiming a success rate above 90% on complex requests. On documents, a 200-page tender file can be assessed locally in about 20 minutes across technical, commercial and pricing dimensions, with scoring rationale for each item.

The weights are live on GitHub, Hugging Face, Gitee, ModelScope and ModelHub with API access, and by the time of the report it had climbed to fourth place on Hugging Face's model trending list. For regulated enterprises with limited compute, the combination of a consumer GPU, domestic training hardware and a domestic framework is becoming a practical route to local agents.