DeepSeek Releases DSpark for Faster LLM Inference

DeepSeek has released DSpark, a speculative decoding framework designed to accelerate large language model inference in production environments. DSpark reuses existing DeepSeek-V4 weights but adds a parallel draft backbone and a tiny sequential head to optimize performance. It incorporates a confidence head and a load-aware scheduler to dynamically adjust token verification based on GPU utilization. In production, DSpark accelerates DeepSeek-V4's per-user generation by 60-85% compared to their previous single-token (MTP-1) baseline. The framework is open-source, including checkpoints and training code.