DeepSeek and Peking University Open-Source DSpark Inference Acceleration Framework

Large language model services handling high volumes of concurrent requests often struggle to balance overall throughput with text-generation speed for individual users because of competition for computing resources and scheduling constraints. DeepSeek and Peking University jointly developed DSpark to improve online inference scheduling and ease performance bottlenecks as popular AI services scale.

DeepSeek and Peking University recently published a paper on DSpark and released the framework as open source. DSpark has also been deployed in DeepSeek’s online service system. Tests showed that it increased client-side text-generation speed by as much as 85% while maintaining the same throughput. Available information did not disclose the exact release date, test hardware configuration or complete benchmark data.