DeepSeek Releases DSpark, a Speculative Decoding Framework That Accelerates DeepSeek-V4 Per-User Generation 60–85% Over MTP-1

DSpark is a serving optimization framework that speeds up text generation from large AI models by combining two different drafting approaches: a fast parallel method and a lightweight sequential component. The system works by having a smaller draft model propose multiple tokens at once, which a larger model then verifies, accepting tokens that match the target distribution without any quality loss. This matters because it reduces latency in production settings where many users are generating text simultaneously, with measured speedups of 60 to 85 percent over previous single-token baselines depending on the workload type. The framework includes a confidence scoring system and load-aware scheduler that adjust how many tokens get verified based on whether the system is busy or idle, allowing it to optimize speed while maintaining output quality.