DeepSeek Deploys DSpark Framework, Boosts DeepSeek-V4 Generation Speed Up to 85%

DeepSeek deployed its speculative sampling acceleration framework DSpark for DeepSeek-V4 production services, following a joint technical report with Peking University. DSpark increases generation speed by 60% to 85% for the Flash version and 57% to 78% for the Pro version without compromising output quality.

DeepSeek open-sourced the full-stack DeepSpec codebase alongside the framework. DeepSpec supports popular open-source models including Qwen3 and Gemma, providing developers with a complete Python toolchain for training custom acceleration modules locally.