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

On September 29, 2025, the DeepSeek-V3.2-Exp model was officially released and open-sourced. 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 listed on the Huawei Cloud Large Model as a Service platform MaaS. For the DeepSeek-V3.2-Exp model, Huawei Cloud continues to deploy it using the large EP parallel solution, based on the sparse Attention structure superimposed to achieve a context parallel strategy for long sequence affinity, while also considering model latency and throughput performance.