On September 29, the DeepSeek-V3.2-Exp model was formally rolled out as an open-source innovation. This particular model stands out as the inaugural one to integrate a fine-grained sparse attention mechanism. This breakthrough substantially boosts the efficiency of both long-text training and inference processes, all while preserving the model's output quality. Owing to the considerable decrease in service costs brought about by this new model, the official API pricing has undergone a corresponding downward adjustment. Developers leveraging the DeepSeek API will now witness a cost reduction exceeding 50%, with the updated pricing structure coming into effect without delay.