Xing4.0 puts 29B agent model on one GPU

China Telecom AI released Xing4.0-29B-A4B, an MoE agent model with 29 billion total and 4 billion active parameters that the company says can run in 15GB of GPU memory with a 256K-token context window, per its release. The model is available through GitHub and Hugging Face, but its 75/100 SWE-bench Verified score is vendor-reported. For builders testing private or local agent prototypes, the useful fact is memory threshold—not benchmark rank: Snapdragon’s local-model analysis shows why parameter count alone is a poor deployment budget.