China Telecom AI Releases Xing4.0-29B-A4B MoE Model Trained on Huawei Ascend NPUs

China Telecom AI has released Xing4.0-29B-A4B, an open-weights Mixture-of-Experts (MoE) language model featuring 29 billion total parameters with 4 billion active per token and a 256K native context window. It is notable for being the first model of its scale trained entirely on Huawei Ascend 910C NPU clusters using the MindSpore framework. The release demonstrates China's growing capability to train competitive frontier AI models natively on domestic Huawei hardware, bypassing reliance on Nvidia GPUs. Its agent-oriented architecture and efficient 4B active parameter budget make it well-suited for complex reasoning tasks and low-cost enterprise deployment. Built on mHC, Multi-head Latent Attention (MLA), and Multi-Token Prediction (MTP) architectures, the model utilizes 64 routed experts and 1 shared expert, activating 4 experts per token. Co-optimizations such as Ascend C mHC fused operators and fine-grained MoE communication optimization improved training throughput by approximately 96% over standard baselines.