On September 29, 2025, the DeepSeek-V3.2-Exp model was officially released and open-sourced. This model incorporates a sparse attention architecture, which effectively reduces computing resource consumption and improves model inference efficiency. The model is now available on Huawei Cloud's MaaS platform. For the DeepSeek-V3.2-Exp model, Huawei Cloud continues to deploy it using the large EP parallelization solution. This strategy leverages the sparse attention structure to achieve long-sequence affinity and contextual parallelism, while balancing model latency and throughput performance.