DeepSeek-AI has published a paper on DSpark, a speculative decoding framework the company says delivers a 2-3x end-to-end throughput improvement over standard autoregressive decoding, according to the paper posted to its DeepSpec GitHub repository.
Speculative decoding works by using a smaller, faster "draft model" to propose several tokens at once, which a larger "target model" then verifies in a single pass, accepting whole chunks when correct rather than generating one token at a time. DSpark's contribution, per the paper, is a hybrid approach that can pair a lightweight dedicated draft model with a self-drafting mode, where the target model itself generates part of the speculative sequence.
The paper reports testing on open-source models including DeepSeek-V2 and Llama 3, with the hybrid design intended to balance draft speed against verification accuracy. Because DSpark supports self-drafting, it can in some configurations skip training a separate draft model, which DeepSeek-AI frames as lowering the effort needed to deploy inference acceleration.
Faster inference lowers the cost of serving LLMs and reduces response latency, which is why speculative decoding has become an active area of research across the industry. DeepSeek-AI's paper positions DSpark as a flexible entry in that space rather than a single fixed technique, adaptable to different model architectures and hardware setups.