Following its $7 billion funding round, Chinese AI lab DeepSeek has introduced DSpark, an open-source speculative decoding framework that accelerates large language model inference by 60% to 85%. DSpark is a specialized engineering layer deployed across DeepSeek-V4 (Flash and Pro) models without modifying core weights or sacrificing output quality. Its core innovation solves suffix decay using semi-autoregressive generation and confidence-scheduled verification, which dynamically calculates draft acceptance probability and real-time GPU load. In multi-domain offline benchmarks, DSpark increased accepted sequence length by 26.7% to 30.9% versus Eagle3, by 16.3% to 18.4% versus DFlash, and sped up user generation by 60–85% (Flash) and 57–78% (Pro) versus DeepSeek’s legacy MTP-1 baseline.