China Telecom unveiled its next-generation Xing4.0-29B-A4B large model on September 17th. It has a total of 29 billion parameters, but only 4 billion are activated. It can naturally handle a context of up to 256K and even expand to 512K. It is the first large model in China that is entirely domestic from the training chip to the inference deployment, and it is specifically optimized for complex engineering tasks.
The most significant feature is the term "fully domestic." The official has completed this process as a closed loop: Ascend chips are used for training, domestic frameworks are adapted, the architecture is self-developed, and the ecosystem is open-sourced — there is no reliance on overseas supply chains, from the underlying computing power to the upper-level toolchain. For the computing power industry, which is often constrained, this approach of holding the entire chain in their own hands is more meaningful than simply increasing parameters.
The performance also stands up to the reputation. In the SuperCLUE evaluation, its agent capabilities scored 93.52 points, ranking third, with a difference of less than one point from the top two models, Qwen. This shows that this "lightweight" player has already caught up closely with the top tier in engineering tasks.
What developers particularly like is its accessible threshold. After 4-bit quantization, the VRAM usage is reduced to 15GB, saving about 75% compared to FP16. This means that a consumer-grade graphics card such as the RTX3090 or RTX4090 with 24GB of memory can run long-context tasks locally — no need to stack servers or rent cloud cards, making it accessible for individuals and small teams.
