DeepSeek DSpark Boosts Generation Speed by 85% in First Post-Funding Paper

DeepSeek has launched a new framework called DSpark, enhancing generation speed by 60-85% without altering the core model architecture. Developed in collaboration with researchers from Peking University, this framework employs a speculative decoding technique to optimize performance. DSpark separates draft generation from target model verification, reducing latency. It utilizes a semi-autoregressive architecture that combines parallel networks with lightweight modules, improving token dependencies within blocks.