DeepSeek and Peking University released DSpark, a speculative decoding framework that accelerates LLM inference by 60% to 85% without compromising output quality. The corresponding models are already available on Hugging Face. This open research approach contrasts with the secretive practices of major US AI labs, highlighting DeepSeek's commitment to transparency and innovation. It enables faster, cheaper LLM inference for a wide range of applications. DSpark introduces semi-autoregressive candidate generation and confidence-scheduled verification to dynamically optimize speculation length and acceptance rate. The framework has been deployed in DeepSeek-V4-Flash and V4-Pro preview, with the full DeepSpec codebase open-sourced on GitHub.
Background
Traditional large language model (LLM) inference generates tokens sequentially, which is slow and limits throughput. Speculative decoding speeds this up by using a fast draft model to predict multiple tokens at once, which the target model then verifies in a single forward pass. DSpark improves on this by generating all candidates in parallel with a backbone network and then refining them with a lightweight sequential module, combined with a confidence-based scheduler that allocates compute to high-likelihood tokens.
References
Discussion
The community widely applauds DeepSeek for its openness and practical innovation, with users noting that the models are already on Hugging Face and delivering strong performance. Some express excitement about potential integration into local inference tools like DwarfStar.