DeepSeek and Peking University jointly open-sourced the DSpark inference framework, which focuses on speculative decoding technology, improving inference speed by 85% and generation speed by 6–8% without replacing model weights.