DSpark

DeepSeek's speculative-decoding drafter. A semi-autoregressive design (parallel backbone + lightweight serial head) paired with confidence-scheduled, load-aware verification. Deployed in the DeepSeek-V4 serving system, it accelerates per-user generation by 60–85% over the MTP-1 baseline at matched throughput. Open-sourced checkpoints ship with DeepSpec, an algorithm-driven training repo for speculative decoding.