# Speculative Decoding

The Efficiency Frontier: DeepSeek and Peking University Reveal 'DSpark' to Slash AI Latency

DeepSeek and Peking University have released DSpark, a speculative decoding framework that boosts AI inference speeds by up to 85%. The framework uses semi-autoregressive structures and dynamic verification to reduce computational waste and has been open-sourced via the DeepSpec training platform.