China Mobile Cloud has introduced a domestic hybrid system for mixed large language model (LLM) inference that combines graphics processing units (GPUs) and neuromorphic technologies. This system was first demonstrated at the Computing Power Conference in Lanfang, Hebei Province, in 2026.
The development of this stack involved China Mobile Cloud, the China Electronics Technology Nanhu Research Institute, Lynxi, Iluvatar CoreX, Tsinghua University, and Peking University. Yu Jian, Technical Director of the Neuromorphic and Optical Laboratory at China Mobile Cloud, emphasized that neuromorphic capabilities are intended to accelerate existing GPU fleets rather than require their complete replacement.
The core concept involves a clear separation of Transformer workload across different chip types. Attention computations remain on domestic GPUs from Iluvatar CoreX, utilizing their aggregate bandwidth. Feed-forward network blocks, including mixture-of-experts (MoE) experts which are critical for latency, are ported to Lynxi's neuromorphic silicon. This silicon is optimized for in-memory computing and large on-chip SRAM, helping to alleviate the pressure created by the 'memory wall' and 'power wall' in single-architecture GPU clusters when serving highly competitive tokens.
A self-developed model compiler, a high-speed interconnect protocol, and a unified inference mechanism are used to ensure synchronization between these two structures, handling task decomposition, joint scheduling, and result aggregation.
DeepSeek V4 Flash was chosen as the test workload. China Mobile Cloud claims that compared to other domestic GPU clusters, this hybrid stack more than doubles inference throughput and energy efficiency while reducing business operating costs by over 40%. Partners attribute these figures to the architecture based on mature domestic processes, emphasizing the system software and workload placement rather than just the race for processes and nodes. They also believe that other operators already using domestic GPUs can replicate this model.
Conference materials also indicate that subsequent deployment goals for this PD and Attention-FFN split include token factories, AI coding, text-to-video conversion, multi-agent interaction, smart manufacturing, and financial and communication security.
The consortium stated plans to continue improving the heterogeneous engine and gradually open parts of the core framework to domestic manufacturers of GPUs, neuromorphic chips, and computing operators. For cloud service buyers focused on cost per token, the debut represents not so much the coronation of a single accelerator, but a test of whether Attention-on-GPU plus MoE FFN-on-neuromorphic can maintain latency, throughput, and utilization after traffic leaves the conference booth. At the time of launch, no independent third-party cluster benchmarks, separate from the partners' DeepSeek V4 Flash settings, had been published; the debut at the Computing Power Conference primarily documents a reproducible recipe of software and silicon, rather than a ready-made capacity SKU.