DeepSeek-V3.2-Exp model officially released and open-sourced

ChainCatcher message, the DeepSeek-V3.2-Exp model is officially released and open-sourced today. The model introduces a sparse Attention architecture, which can effectively reduce computational resource consumption and improve model inference efficiency. Currently, the model has been officially launched on Huawei Cloud's Model as a Service platform (MaaS). For the DeepSeek-V3.2-Exp model, Huawei Cloud continues to use the large EP parallel deployment scheme, implementing a long-sequence affinity context parallel strategy based on the sparse Attention structure, while also considering model latency and throughput performance.