China Telecom and Huawei Release Open-Source Xing4.0-29B Code Agent Model
China Telecom and Huawei officially released Xing4.0-29B-A4B, an open-source Mixture-of-Experts (MoE) code agent model featuring 29 billion total parameters with 4 billion active parameters per token. It marks China's first tens-of-billions parameter model trained natively on Huawei's Ascend Atlas 900 A3 liquid-cooled SuperPoD supernodes using the MindSpore framework. This launch highlights significant progress in China's domestic AI ecosystem, demonstrating high-performance model training fully independent of NVIDIA hardware and Western software frameworks. Open-sourcing the model across platforms like GitHub and HuggingFace provides developers with a lightweight yet powerful agent for code development, task planning, and long-context analysis. Through joint hardware-software co-optimization—including Multi-head Latent Attention (MLA), mHC fused operators, and fine-grained MoE (64 routes / Top4 active)—training throughput on the Ascend cluster improved by 96%. The model natively supports context windows scaling up to 256K tokens and utilizes multi-expert reinforcement learning for multi-step agent reasoning and execution.
## BACKGROUND
A Mixture-of-Experts (MoE) model routes inputs to specialized sub-networks, activating only a small subset of parameters during inference to maintain high speed and lower execution cost. Multi-head Latent Attention (MLA) is an architectural technique that compresses Key-Value (KV) cache memory, making long-context processing far more compute-efficient. Huawei's Ascend Atlas 900 A3 SuperPoD is a specialized domestic AI computing cluster capable of interconnecting hundreds of Ascend chips into a unified supernode.