The system was unveiled during the 2026 China Computing Power Conference, held in Langfang, Hebei, from September 11 to 13. It is the result of a collaboration between Mobile Cloud, China Electronics Technology Group Nanhu Research Institute, Beijing Lingxi Technology, Shanghai Iluvatar CoreX, Tsinghua University, and Peking University.
Testing with the Deepseek V4 model indicates that the heterogeneous architecture improves cost-effectiveness by more than double compared to standard domestic GPU clusters. Additionally, the system reduces operational costs by over 40 percent.
The architecture utilizes a divide-and-conquer strategy, splitting large models into PD and AF components. Attention calculations are handled by domestic GPUs, while latency-sensitive FFN modules are processed by brain-inspired chips that leverage on-chip SRAM.
The system integrates proprietary model compilers, high-speed interconnect protocols, and a unified inference engine. These components manage task decomposition, scheduling, and result aggregation across the different chip types.
The solution is designed for real-time applications such as AI code generation, multi-agent collaboration, and smart manufacturing. It also provides a fully localized inference foundation for security-sensitive sectors, including finance, security, and telecommunications.
