SYLIN

DSpark

by DeepSeek · Technique · Text & code

Multi-token-per-step decoding component included in V4.1 Flash

Evidence

  1. Sep 7
    DSpark gave 1.15–1.63x on Qwen3-8B on AMD GPUs versus 1.08–1.27x for DFlash
  2. Aug 16
    Speculative decoding technique added to DeepSeek V4 Pro 0813
  3. Aug 2
    Lucebox's DSpark drafter uses DeepSeek V4 Flash's built-in multi-token-prediction heads, with no separate draft model needed